awesome-free-ai-coding · every free model · every provider · filter the list
257 services whose free tier the list could not find or could not verify on the date checked. Nothing here is disqualified — domains rejected for cause are in blocklist.yaml — and each verdict expires after 90 days and is asked again. The records live in watchlist.yaml.
2026-11-01 — The legacy Meta Llama API portal at llama.developer.meta.com redirects to dev.meta.ai; both names are retained in this single decision. Current signup and billing help require adding a card before requests. Standard and Contributor API modes have positive token prices; Contributor mode permits training on submitted content. Muse Code uses the same billing account, with metered keys or paid monthly subscriptions. The previously observed $20 welcome-credit message does not establish a current no-card grant. Downloadable Llama and Muse model weights are separate from hosted inference. Primary sources: dev.meta.ai/help/accounts-and-login/sign-up, dev.meta.ai/help/billing/set-up-billing, dev.meta.ai/docs/pricing-rate-limits.md and dev.meta.ai/docs/muse-code/subscriptions.md. No account was created to check introductory-credit activation. Reopens if: Current primary terms establish an obtainable no-card hosted API or Muse Code grant, or a recurring free-model allowance. Review all modes independently of paid token prices; retain the legacy portal aliases.2026-10-31 — The public announcements and linked campaign page describe a $1 one-day trial and a 200M-token benefit for a $10 recharge during October 15–21, 2026; that dated campaign has ended. A separate October 30 first-deposit matching bonus expires 30 days after crediting. Current public model prices and API billing docs do not establish a continuing free coding allowance or a new no-payment trial outside that campaign. The signup console and API require an account; activation was not tested. Reopens if: Current primary terms or an existing authorized account establish an obtainable no-card API trial, a new free-model campaign, or recurring free API usage. Re-read the campaign and announcements; a paid price does not negate a separately documented free mode.2026-10-31 — The current platform offers $10 promotional credit after a $1 card top-up, once per person and card. This is a one-time card-required bonus. The homepage and public API pricing meter the text models; no recurring free-model allowance is advertised. Terms mention sign-up credit generically but do not establish a current obtainable no-card grant. Public status disables check-in. No new account, key or payment was created to test activation. Reopens if: Current primary offer or an existing authorized account establishes an obtainable no-card signup grant or recurring free model quota. Review all access modes independently of positive per-token prices and the card-funded promotional bonus.2026-10-30 — Deyin Code’s current pricing page (https://deyin.ai/pricing) advertises an Openference Free plan with 100 requests every five hours and 350 requests a week, no card required. The current public plans catalog (https://api.openference.com/api/public/plans) instead returns only paid Minion and Boss subscriptions. That catalog does not establish whether a separate default Free entitlement exists. The advertised no-card API or bundled CLI access remains unconfirmed; the older Promo plan and its prices no longer describe the current catalog. Reopens if: Current primary documentation or a no-card account establishes the advertised Free entitlement, its eligible API or Deyin CLI access, and its limits and payment conditions. A separate paid plan does not disqualify a demonstrated Free mode.2026-10-26 — The current Amazon Q Developer end-of-support announcement closes new Free accounts and new subscriptions from May 15, 2026. Existing subscriptions can add users and existing IDE users retain access until April 30, 2027. This is a new-user eligibility exclusion, not a claim that all existing users lost the product in May. Source: https://aws.amazon.com/blogs/devops/amazon-q-developer-end-of-support-announcement/. Reopens if: AWS reopens Free account creation or documents a new bundled free coding offer reachable by new users. Check existing-user access independently.2026-10-22 — Current direct API documentation prices every listed model and publishes no automatic signup credit or no-payment API allowance. The quickstart requires an account and API key and points to the wallet for top-ups. The documented free chat app and Trinity routes on OpenRouter do not establish a free direct API offer. Reopens if: Arcee publishes directly usable API credits or a free model allowance without an application or recurring paid subscription.2026-10-22 — The Hermes catalog and fresh ordinary, streaming and invalid-key calls return 401 public_beta_closed: the public beta is closed, an API key is required and key registration is closed. The website still describes the keyless endpoint, contradicting the live service. The published row and its history are retained in Archive. Reopens if: The public endpoint reopens or the operator documents an obtainable free API key and current callable chat model.2026-10-15 — Current onboarding (https://blueclaw.network/ and https://blueclaw.network/quickstart) requires buying prepaid credits before API use and no longer advertises the previous $5 welcome grant. The terms (https://blueclaw.network/terms) describe promotional credits generically but establish no current signup amount or eligibility. The public portal shows email sign-in without a free-grant claim. An account-specific grant has not been measured. The archived row preserves the previously published offer. Reopens if: A current primary offer establishes a usable free model quota, recurring credit or signup grant. Check exact eligibility and any initial-payment condition before restoring the existing row.2026-10-15 — The current FAQ (https://hyper.charm.land/faq, also served as https://hyper.charm.land/prose/faq) explicitly retires the monthly free Hypercredit allowance. Remaining access requires paid Shred/Business subscriptions or prepaid credits starting at $5. Older billing docs (https://hyper.charm.land/docs/plans-billing.html) and the Crush README still describe the previous free plan; that conflicts with the current retirement notice. The archived row preserves the previous offer. Reopens if: Current primary terms restore bundled free usage or a usable free grant. Reconcile the current FAQ, billing docs and account eligibility before restoring the existing row.2026-10-15 — The current pricing page gives Free accounts 0 requests/week and requires a paid plan for API inference, starting at $19/month (https://yolo-auto.com/pricing). The model catalog labels inference access as paid (https://yolo-auto.com/models). No-card account creation does not include a usable inference allowance. The previously listed offer and its history remain in the archived row. Reopens if: Primary terms introduce a positive free API allowance or usable trial compute without a recurring paid subscription. Restore the existing row after checking current activation and model conditions.2026-10-14 — The developer page advertises a no-card Free API allowance of 1,000 zero-cost-model requests per calendar month, shared across keys and models, with 20 requests/minute. Terms instead require paid membership for service access. Developer pricing does not resolve that conflict or establish an independently usable tool-calling lane. No signed-in eligibility or completion was established. Reopens if: Terms and developer pricing agree that API Free needs no recurring paid membership, and exact free model/tool and supply conditions can be verified.2026-10-14 — The primary Free plan is $0/month with $5/workspace/month in shared usage credits. The official OpenAI guide names moondream3.1-9b-a2b and moondream3-preview, multi-turn text/image chat and SSE. Stop sequences and tool/function calling are explicitly unsupported, so an OpenAI-compatible label does not establish a coding-agent tool loop. Remote image URLs are rejected; only base64 data URLs work. Local Photon usage is separate. Keyed activation and the initial credit balance have not been established. Reopens if: The official cloud API supports the coding-agent tool loop, with current Free activation/payment conditions and an authorized representative completion.2026-10-14 — The free-routing-only verdict is outdated. The active public free-package and pricing APIs name four generative zero-price aliases beside the OrcaVerify classifier. Baseline access is 10 requests/minute and 50/day without a top-up; HY4 Preview has a separate 50/minute and 250/day baseline. Established GitHub linkage gates free access. Docs describe an unpublished prompt cap and separate access-denied, quota and capacity failures; free IDs do not fall back to paid IDs. Terms name Continuum AI Pte. Ltd. and commercial upstream providers, but actual free-lane supply and eligible-account generation remain unproved. Classifier access does not add a coding model. Reopens if: An authorized established-GitHub account confirms the generative free lane, its serving route and current prompt/plan limits, while primary free-package and pricing APIs still agree.2026-10-14 — Docs distinguish permanent $0 Micro/Lite aliases from a seven-day trial. The catalog describes changeable upstreams and fallback chains conditional on health, plan and budget; aliases cannot be copied as independently free upstream families. The shared ceiling is 30 requests/minute and five concurrent requests; token allowance is unpublished. The Terms link returned the home page. Activation/payment, free supply and model-specific tool behavior remain unresolved. Reopens if: Readable Terms and plan documents establish the permanent free API and its supply/activation conditions; an authorized tool-loop request verifies the free alias and fallback constraints.2026-10-14 — Primary pricing now lists Ternary Bonsai 27B at zero input and output price, replacing the old all-paid verdict. Official billing requires an initial $15 credit purchase and offers no free trial. The competitor claim of 98 free Together models is unsupported by primary pricing. Exact independent free access after activation has not been established; no purchase was made. An initial payment is an access condition, not automatic disqualification. Reopens if: Primary sources establish the exact zero-priced API ID and activation conditions, and an authorized funded account confirms the independent free lane; or the activation payment is removed.2026-10-13 — The current free price table (https://beatapi.io/pricing) advertises GPT-6.1 Sol, DeepSeek V4 and V4.1 Flash, GLM 5.3 and 5.3 Flash, and Gemini 3.8 Flash at zero input and output cost under separate free IDs. The product page says no card or top-up is required (https://beatapi.io/gpt-6-1-sol-api); the free-route docs support Responses and Chat Completions (https://docs.beatapi.io/text-api/gpt-6-1-sol-free). Free calls share one successful request per minute per account before the first top-up and at most ten afterward (https://docs.beatapi.io/faq/account-billing). Daily allowance and continued availability are not guaranteed. The seventh free item, JEV, returns typed decisions rather than generated text (https://docs.beatapi.io/decisions). Public terms and routing information (https://beatapi.io/terms-of-service and https://beatapi.io/llms.txt) do not establish self-hosted inference or named authorized upstream capacity for the free text routes. That does not establish unauthorized proxying. The model catalog and fresh Chat Completions and Responses requests without a key return 401; authenticated zero-balance completion and a tool round trip remain unverified. Reopens if: BeatAPI establishes self-hosted inference or named authorized upstream capacity for its free text routes, and a no-card account demonstrates free completion and a valid tool-result round trip with the advertised ID and payment conditions. Re-read current free prices, daily allowance and promotion terms before inclusion; any documented initial payment is an access condition, not an automatic rejection.2026-10-13 — The site and public catalog are live. The documented :free models cost nothing and share 10 RPM and 100 RPD per account (https://docs.naga.ac/build/rate-limits). Routing documentation says requests go to changing third-party providers, whose identity is never disclosed in responses and cannot be selected or excluded (https://docs.naga.ac/build/routing). It does not establish self-hosted inference or named authorized upstream capacity. Matching model IDs or lower prices alone do not establish unauthorized proxying. The terms require purchasing credits while the pricing and billing docs advertise zero-cost models; fresh missing-key and invalid-key requests both return 401, so no zero-balance completion is established. Free-model data collection for training by NagaAI and its partners is always enabled (https://naga.ac/legal/privacy). Reopens if: NagaAI discloses self-hosted inference or named authorized upstream providers and documents the free lane’s payment and eligibility conditions, including any minimum initial payment. A documented one-time deposit is a condition, not an automatic rejection; model-ID overlap alone is not a rejection reason.2026-10-08 — The current official pricing table at https://huggingface.co/docs/inference-providers/pricing lists Free Users with no monthly credits. Inference requires purchased credits or a paid PRO, Team or Enterprise subscription; BYOK offers no bundled free usage. The former $0.10/month allowance is no longer advertised. Reopens if: Hugging Face publishes bundled free inference usage or a no-card trial available to free accounts, without buying credits or a subscription.2026-10-05 — Free chat and streaming are real, with no key, wallet or payment header: https://blockrun.ai/free documents automatic substitution and the response names the model that answered. The seven free chat rows at https://blockrun.ai/api/v1/models do not establish seven callable models: five were marked unavailable, then six on a fresh read. Llama 3.2 11B Vision answered under its own ID but returned non-JSON tool arguments; the tool-result turn failed with an upstream parse error. Laguna XS 2.1 completed a tool read and answer through Nemotron 3 Super 120B instead, which the catalog does not list, and Laguna was subsequently marked unavailable too. The published free 1M contexts belong to unavailable rows. https://blockrun.ai/docs/api-reference/chat-completions documents the canonical /api/v1 route and /v1 alias; both were tested. Paid-model control correctly returned an x402 payment challenge. The offer is legitimate, but no catalog-listed available free ID demonstrated both its advertised identity and a valid tool round trip for the configs. Reopens if: The catalog exposes an available zero-priced coding lane that completes a tool call with JSON arguments and consumes its result, with the actual served model matching the catalog or an explicitly cataloged free router ID. Recheck the catalog and live calls when those routes are repaired.2026-09-30 — Its no-card free tier is speech. The pricing page answers “Can I sign up for free?” with “You can create an account and start transcribing immediately with no credit card required. The free tier includes up to 185 hours of pre-recorded transcription and up to 333 hours of streaming transcription.” Its LLM Gateway is OpenAI-compatible, at https://llm-gateway.assemblyai.com/v1, “One API for Claude, GPT, Gemini, Qwen, Kimi, and more”, but the gateway’s rate-limit page gives the Free account type “Not available” and a Paid one 30 requests a minute per model (assemblyai.com/pricing, docs/llm-gateway/rate-limits, read 2026-09-30). pacocartones/free-llm-api-hub lists the signup credit beside the gateway as though the one reached the other. Reopens if: The LLM Gateway’s rate-limit page gives free accounts a limit, or an AssemblyAI page says the free tier reaches the gateway.2026-09-30 — A RAG platform: its pricing page offers “Claim your $25 in free credits” on the on-demand plan, spent on datastores, agents and component APIs for parsing, reranking and grounded generation. Text comes out of its agents and out of Generate, a POST to https://api.contextual.ai/v1/generate for its Grounded Language Model, “an LLM engineered specifically to prioritize faithfulness to in-context retrievals”, capped at 32,000 tokens a request; nothing in its documentation index (docs.contextual.ai/llms.txt, read 2026-09-30) is a chat-completions or messages endpoint a coding agent could be pointed at, so the credit buys nothing this list carries. Reopens if: Contextual AI documents an OpenAI- or Anthropic-compatible chat endpoint that the free credit reaches.2026-09-30 — Its Model APIs are OpenAI-compatible, take Claude Code through ANTHROPIC_BASE_URL, and “Users get up to 40 million free tokens to start” at 15 requests a minute on the Free plan (docs/platform/inference/model-apis). The account is free — “no credit card required” to create it, a non-virtual phone number to verify it — and “Eligible free plan users receive complimentary credits”, which “can be used for compute, inference, and storage” and are “promotional” (the getting-started, FAQ and billing pages, read 2026-09-30). No page says the credits reach an account without a card, and the console offers them beside a card: its billing panel pairs “Claim your … free credits” with an “Add card” button, and its sign-up promises “up to 30 free credits”. The 44 models of the keyless catalog at lightning.ai/api/v1/models are all priced, Claude Opus 4.8 and GPT-6 Astra among them, so the credit is the whole offer — the SiliconFlow record, the card question open. Reopens if: A Lightning page says the free credits reach an account without a card, or a model in its catalog is priced at zero.2026-09-30 — Alibaba’s developer platform for Qwen and third-party models, run by Intelligent Cloud Computing (Singapore), OpenAI-compatible at https://maas.qwencloudapi.com/compatible-mode/v1 with Claude Code and Codex guides. Its homepage says “Most models offer free API credits”, and its free-quota page gives each model a quota valid 90 days from activation, but the way in is “Sign up for QwenCloud and complete your payment information to activate your free quota” (docs.qwencloud.com/resources/free-quota, read 2026-09-30) — a credit behind payment details, which CONTRIBUTING does not take. Alibaba Cloud Model Studio’s international free quota is a row. Reopens if: A QwenCloud page grants the free quota before any payment information is given.2026-09-30 — Known only from three repositories an account opened on 2026-06-03 — willianhecaiyan/claude-no-creditcard-demo, TOKENTOWORLD and fix-openai-429-demo — which sell it as Claude 3, GPT-4o, DeepSeek, Kimi and Qwen for developers “lacking US credit card”, with a “Stable multi-source balance to avoid interface blocking”, wholesale and affiliate terms, and “trial tokens” of no stated size. Nothing answers to read: tokentoworld.com and www refuse HTTPS here and from a GitHub runner on 2026-09-30, and plain HTTP returns nginx’s 404 page. A relay whose offer cannot be read cannot be verified to offer anything, the rule that settled Clarifai, Coze and Morph, and the ad does not say where the Claude it sells comes from. Reopens if: tokentoworld.com serves its terms over HTTPS, stating a trial spent before any payment and where its Claude supply comes from.2026-09-29 — The API prices every model — MiniMax-M3 at $0.30 / $1.20 per million tokens after a “Permanent 50% off”, M2.7 and the older M2 line at $0.3 / $1.2 — and sells Token Plans from “$22 /month” and Credits at “1,000 credits = $1” (platform.minimax.io/docs/guides/pricing-paygo and pricing-token-plan, read 2026-09-29). One tutorial there says “New users without subscription can try the free quota from Pay-as-you-go”, and the rate-limits page speaks of a free tier, but no page states the quota’s size or how it is reached; an amount the vendor does not print is not a quota this list can carry. MiniMax Code, the coding agent on agent.minimax.io, is a row of its own. Reopens if: A signup grant or free lane with a figure on a vendor page.2026-09-29 — An AI workspace over other vendors’ models whose Free plan is chat — “15 messages per day” with a “Limited model selection”, “no card required” (zerotwo.ai/pricing, read 2026-09-29). Its coding agent, ZeroCode, a desktop agent for local repositories, is “Not included” on Free in the plan table, which gives it from Plus at $14.99 a month (the docs, from Pro); the code-agent page’s “Free to try” meets that table, and the table is the specific statement. Chat messages in a workspace are not model use a coding agent or an endpoint can spend. Reopens if: ZeroCode comes with the Free plan, or a free allowance a coding agent or an endpoint can spend.2026-09-27 — Its get-started guide asks for money first — “The minimum top-up amount is $25” (atlascloud.ai/docs/en/get-started, read 2026-09-27) — while its model pages say “New users receive free credits to explore the platform and test models”, of no stated size and with no stated way to reach them before paying: the SiliconFlow record. Reopens if: A page states the free credit’s size and that it is spent before any payment step.2026-09-27 — Paid only: subscriptions or “Pay just 1¢ per credit” on codebuff.com/pricing (read 2026-09-27). The same company’s free product is Freebuff, a row. Reopens if: codebuff.com publishes a free allowance of its own.2026-09-27 — Zhipu’s IDE plugin says “CodeGeeX is available for free to all developers as a plugin product on your IDEs” in its VS Code and JetBrains marketplace descriptions, but its last VS Code release is 2025-09-01 and its last JetBrains one 2025-12-11, codegeex.cn serves a script shell, and nothing a probe reads would change the day the free service stopped answering — the descriptions are a year old (read 2026-09-27). Zhipu’s current coding agent is the ZCode row. Reopens if: A release after 2025-12-11, or a page the client reads that states the free plugin plan today.2026-09-27 — Free and non-commercial (“Free Edition accounts may not be used for commercial purposes”), in a “serverless-only, quota-limited environment” where you can “experiment with foundation models”, but its limits page names no model and no quota for them — “Certain models not available” — and the endpoint is each workspace’s own (docs.databricks.com/aws/en/getting-started/free-edition and free-edition-limitations, read 2026-09-27). A reader cannot tell what they would be calling. Reopens if: Databricks documents which foundation models a Free Edition workspace can query and at what limits.2026-09-27 — A VS Code, JetBrains, Visual Studio and Vim coding assistant on Fitten’s own model with no bring-your-own-key, whose individual use is free by its home page’s pricing section and its FAQ — both rendered by script, which no probe can read: code.fittentech.com serves the project’s client a 1 KB shell, and neither its VS Code nor its JetBrains marketplace description mentions a free plan (read 2026-09-27). Reopens if: A page the project’s client can read — a static pricing or docs page, a marketplace description — states the free individual plan.2026-09-27 — Its pricing page still offers a Hobby plan that “Includes 50 premium model requests per month 1000 autocompletions 100 basic model requests” and calls the VS Code agent “free forever with limited requests”, but the extension’s last release is 2025-07-09, fourteen months before this read, and a page that stands still is no sign the service answers — the probe would vouch for a product that may be gone (read 2026-09-27). Reopens if: A release after 2025-07-09, or other evidence that the agent still serves its free requests.2026-09-27 — An open-source agent that brings no model of its own: “Works with 15+ providers … Use API keys or your existing Claude, ChatGPT, or Gemini subscriptions via ACP” (goose-docs.ai), and its free options are other vendors’ free tiers — “not all supported LLM Providers provide a free tier” (goose-docs.ai/docs/getting-started/providers/, read 2026-09-27). BYOK only. Reopens if: goose ships a hosted model lane of its own with a free allowance.2026-09-27 — A Vietnamese gateway whose own “Kira” models, by its FAQ, combine in-house open-source models with “OpenAI (GPT 5.5), Claude (Claude 4.8) và Gemini (Gemini 3 trở lên)”, its packages apply “chỉ áp dụng cho model Kira” (only to the kira- models), and partner models are paid from a VND wallet (kiraai.vn/faq/, read 2026-09-27). No chat model is free in its own right. Reopens if: A kira- chat model is priced at zero, with the models behind it named.2026-09-27 — A one-person gateway over other vendors’ free tiers: “The free-route models run on the free tier each provider offers; onomeo puts them behind one key”, and “Big models such as Claude and GPT are paid for by onomeo, within a shared daily budget” (onomeo.com/models, read 2026-09-27). The free tiers it routes to are rows here in their own right; a gateway’s free plan over them is a watchlist verdict, never a row. Reopens if: Onomeo serves models of its own, or buys the capacity it gives away, and says so.2026-09-27 — “15+ providers, hundreds of models … Authenticate via API keys or OAuth” (pi.dev, read 2026-09-27): bring your own key or subscription, no bundled model usage. Reopens if: Pi bundles a free model allowance of its own.2026-09-27 — “Is Pieces free to use? No. … New users start with a 7-day free trial on any plan”, and “Starting the 7-day Pro trial takes a card on file” (pieces.app/pricing, read 2026-09-27) — a trial behind a card, which CONTRIBUTING excludes. Reopens if: A plan or trial without a card on file.2026-09-27 — “As of July 23, 2025, the following plans will no longer be available: Cody Free, Cody Pro” (Sourcegraph’s blog, read 2026-09-27); Cody is an Enterprise product now. Lists still carry the free tier. Reopens if: Sourcegraph offers an individual plan with free Cody usage again.2026-09-27 — A reseller of other vendors’ models whose free offer is a one-off trial: “Start free with 100 requests — no credit card, no payment required” (yingsuan.top/payment.html, read 2026-09-27) — a hundred calls, then paid. Reopens if: A recurring free allowance, or free models it serves itself.2026-09-26 — The international site prices every model and offers “postpaid billing—plus $1 in free credits to get you started!”, a one-off grant, and no page read says whether the dollar can be spent before a payment method is added (2026-09-18). Read again 2026-09-26: all 87 models on its models page carry a price (from $0.003 per million input tokens), the console’s sign-in says “Get started now and enjoy $1 credit to explore our platform!”, and still no page says whether a card is asked first. The Chinese site, siliconflow.cn, is a row: its free chat models open to an account verified with Chinese, Hong Kong, Macau or Taiwan papers, which is rank and not a reason to leave it off, as for Tencent TokenHub. Reopens if: A siliconflow.com page says the $1 needs no payment method, or a model there is priced at zero.2026-09-26 — The international site’s New User Free Trial Package gives each language model a million tokens valid a year, in a campaign to 2026-12-31, but “Each time you claim a free resource package for a model or enable postpaid service for a model, USD 1 will be automatically frozen in your account. If your account balance or credit limit is insufficient, the claim or enablement will fail” — a funded account, model by model (read 2026-09-26 in a browser). Its docs answer the project’s client and the runners alike with EdgeOne’s JavaScript challenge, so no probe could read a row back. The Chinese site, cloud.tencent.com, is a row (tencent-tokenhub) for accounts verified with mainland, Hong Kong, Macau or Taiwan papers. Reopens if: The package can be claimed without a balance, or the offer is published on a page a probe can read.2026-09-23 — Listed from 2026-08-06 to 2026-09-23 as a free tier that applies while no payment method is linked, and taken off because a new account gets no free usage. The console’s plans page, cloud.sambanova.ai/plans, read “Start exploring SambaNova Cloud with free API credits. No credit card required.” in every Wayback capture up to 2026-06-20; by 2026-08-03 it read “Start exploring SambaNova Cloud. Add a payment method and purchase credits to run your first requests.”, as it still does, and on 2026-08-12 a user on SambaNova’s forum asked why, for a friend’s new account, “a payment method is needed to use it” — unanswered. The docs’ rate-limit page still describes a Free Tier “Applied when there is no payment method linked with your account”, which is the page the row’s probe read, so the probe passed through the change; the tier may hold for accounts that had free credits before it. Reopens if: The plans page offers free credits or free usage without a payment method again.2026-09-22 — The Turkish Presidency of Defence Industries’ AI platform serves open models through an OpenAI-compatible API on H200s in Turkey, and its requests pass “kota ve kredi kontrollerinden” (through quota and credit checks), the credits earned by labelling data or spent on inference. The site has one way in, “e-Devlet ile Giriş Yap” (sign in with e-Devlet), through SSB’s identity server on its EDEVLET realm — Turkey’s state portal, which takes a Turkish identity number: a national identity wall. LLMTR’s evren/* ids run on the caller’s own EVREN key and are ignored on that row for the same reason. Reopens if: EVREN takes sign-ins that need no e-Devlet account.2026-09-21 — A gateway run by the French association Skyclouds whose Free plan gives “€20 Monthly usage credits”, “Renewed every month · No credit card required”. Its catalog serves most models twice: “distribué par l’endpoint Anthropic T2+ configuré par ACLIDE” at list price, and “est disponible via l’API unifiée ACLIDE” below it — and the second route carries all 36 zero-priced ids, Mistral’s whole line, MiniMax M2.x, DeepSeek V4 and Qwen ‘:free’ ids, other vendors’ free quotas. Who runs that route is never named: the privacy page says only “Administration may select another secure AI endpoint operator”. The xKiro case. Reopens if: The catalog names the operator behind “l’API unifiée ACLIDE”, and it is the vendors’ own capacity or sold to ACLIDE for this.2026-09-21 — “Token Plan is an AI model subscription service offered by Alibaba Cloud Model Studio”, $6 to $68 a month; models.dev lists its rows at cost 0 because the usage is the subscription’s. The free quota Model Studio does give is a row of its own. Reopens if: A Token Plan tier priced at zero.2026-09-21 — Unreadable: blazeapi.org renders 30 characters without JavaScript and docs.blazeapi.org does not resolve. What is known of the offer is a router’s label for it, “daily free tokens; Discord verification”. Reopens if: Docs or a pricing page readable without JavaScript that state the free tokens.2026-09-21 — A pay-as-you-go coding gateway (GLM-5.3, DeepSeek V4 Pro) whose home page FAQ says “Top up credit from $2” and names no free tier. The only statement of one is a July blog post, “CloudCode.ONE now has a free tier”, published with its figures never filled in — “[X tokens / $Y worth of credit] per [day/month]” and “[The quota resets every day/month.]” — so the offer has no quota a reader could plan against. OmniRoute’s config calls two GLM Flash ids there “published free model aliases”. Reopens if: A CloudCode.ONE page states what the free tier gives, in numbers.2026-09-21 — “100 free daily requests, auto-replenished” across “50+ frontier and open-weight AI models”, priced “always below the provider’s list price cap” — a reseller that names neither the models the free requests reach nor where it buys them; its site renders client-side and docs.clod.io does not resolve. Reopens if: A readable page naming the free models and the supply behind them.2026-09-21 — Plans start at Starter, “$19 /month 4M credits/month”, and the home page, read whole from a GitHub runner, never says “free”. Reopens if: A free plan or no-card trial credit on cosine.sh/pricing.2026-09-21 — The free lane is one router id, dgridai/free, which “aggregates multiple sources of free inference capacity behind a single routing layer”; the model behind it “can change from request to request” and no page names it or its sources. 10 requests a minute and 100 a day per account, 20 and 1,000 after $5 of top-ups. Free capacity gathered from other sources is the NaraRouter verdict above: a gateway re-serving free lanes it does not own is not the vendor of the offer. Reopens if: DGrid names the models and the sources behind dgridai/free, and they are its own capacity or sold to it for this.2026-09-21 — Says what it is: “One free key routes across the free tiers of many providers with automatic failover” — “Free tiers first, trial credits next, keyless ch.at as the always-on last resort”, over Groq, Mistral, NVIDIA NIM, Cohere, Google AI Studio, OpenRouter and Cloudflare. The offer is other vendors’ free tiers under one key, the NaraRouter verdict. Reopens if: Free capacity of its own, or the vendors’ consent to resell theirs.2026-09-21 — One of the Discord-membership relays sources.yaml records under the YoannDev90 and zukixa lists, where nothing is a vendor’s own offer; its site renders 639 characters without JavaScript, and a router’s label for it is all that describes the free part, “shared weekly credits”. Reopens if: A readable page that names the free credits and where they are served from.2026-09-21 — The product page still sells “Gemini Code Assist for individuals at no cost” with last year’s copy — “the power of Gemini 2.5 to your IDE at no cost (with Gemini 3 coming soon)” — but Google’s developers blog of 2026-05-19 ended it: “On June 18, 2026, Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge using Gemini Code Assist for individuals.” The free individual plan moved to Antigravity, which is a row. Reopens if: Google’s docs or blog say the Code Assist IDE extensions serve individual free users again.2026-09-21 — A New API relay run by 新疆幻城网安科技有限责任公司 — its own FAQ opens with “中转站如何使用？” (how do I use the relay station?) — whose free-user group serves other vendors’ models from sources it does not name. On 2026-09-06 the operator warned that the free group’s glm-5.3 “是第三方低价服务商非官方服务商 模型真实度有待考证”: it comes from a third-party low-price, unofficial provider, and whether it is the real model is unverified. A sign-up balance and invite rewards spent on that is not a vendor’s own offer. Reopens if: The free group serves the operator’s own deployment of a named model, with its quota on a page.2026-09-21 — A crypto-funded gateway whose pages disagree about the free offer. The home page grants “$10 free credit on signup” — “$2.000000 instant, another $8.000000 released on request” by support — and greets a visitor with “Your 0 daily claim is waiting”, while the about page promises “1M free Claude Opus 5 tokens claimed daily” and the terms “the 1M daily Claude Opus 5 claim”. No page names the operator: the contacts are a WhatsApp number and a Telegram handle, and the privacy page says only that “we route requests to upstream providers”. Reopens if: A named operator, and one statement of what a new account gets that all of its pages agree on.2026-09-21 — A subscription gateway run by Eqlyx, LLC whose $0 Basic plan gives “∞ requests/day for most of free models” — “a rate limit of up to 100 requests per hour applies to free models”, its FAQ adds — and 50 premium requests a day. The free models are about forty ‘:free’ variants of other vendors’ models (DeepSeek, GLM-5.2 and GLM-5.3-Flash, Kimi, MiniMax, Mistral Large 3, Gemma 4, gpt-oss) and no page says where they are served from: the xKiro case above, unlisted until that is known. Reopens if: LiteRouter names the providers behind its :free ids, and they sell or give it that capacity.2026-09-21 — A roleplay-first API whose free daily tier is other providers’ capacity: “Composite runs a free daily tier of open-weight models, including Llama 3.3 70B, DeepSeek V4 Flash, GLM 5.2, Kimi K2.6”, and “For non LucidityAI models, OpenRouter is the upstream provider”. The same page says both that “connecting a Discord account unlocks the free tier” and that “No Discord account, credit card, or invite is needed”. Reopens if: The free tier runs on Lucidity’s own infrastructure, with its models and quota stated once.2026-09-21 — The Token Plan is a subscription — “Token Plan 通过订阅 Key 提供套餐内 Token Plan 用量额度” (the plan’s quota, through a subscription key), bought as a personal subscription or credits — so its zero-cost rows on models.dev are included usage, not a free tier. Reopens if: A free quota on MiniMax’s China platform that needs no subscription.2026-09-21 — Bring-your-own-key now. refact.ai/pricing is a 404, the main repository (smallcloudai/refact) was archived after 2026-05-30, and docs.refact.ai redirects to a fork’s wiki whose setup page says “Refact is designed around BYOK rather than bundled cloud accounts”. Reopens if: Refact bundles model usage again — a plan with included credits.2026-09-21 — “Start for free — no credit card required”, but what the free Starter plan spends its AI on is Replit’s own hosted workspace — the case CONTRIBUTING excludes, “hosted app builders whose free tokens are only spendable inside their own workspace”. Reopens if: Free model access usable from a client or an endpoint outside Replit’s workspace.2026-09-21 — A flat-rate “Token Plan” gateway with a Free plan, but no readable page states its quota or its models: “on the Free tier it draws only from free-eligible models”, unnamed, and “Router9 fronts multiple providers and routes each request for the best value”, also unnamed. The pricing page redirects to a login. Reopens if: The Free plan’s quota and models on a public page, and the providers behind them named.2026-09-21 — Unreadable: every page answers a bot challenge (“enable javascript and cookies to continue”), read directly and from a GitHub runner alike. The only statement of the offer is a router’s label for it, “daily free-model quota”. Reopens if: A page readable without a browser challenge that states the free quota.2026-09-21 — “The indie LLM router that tries free models first” — “waterfall routing from free to cheap to premium”. A router serves nothing itself, so its 131 free models are other providers’ free lanes re-served under its own key: the NaraRouter verdict. Reopens if: Free capacity of its own, named, with a quota.2026-09-21 — The $0 Discover plan carries “250 AI Credits”, each “approximately $0.01 of model usage”, spent inside “an AI creation platform to build websites, AI apps, dashboards, internal tools, and mobile apps” — an app builder, in the cloud or as its desktop app, not model access a developer can aim at code they already have. Reopens if: The credits become usable from other clients, or Workshop Desktop works as a coding agent on an existing project.2026-09-18 — Nothing free is left. b.ai itself still renders the word BAI and nothing else, but docs.b.ai is server-rendered and its sitemap lists the LLM service’s pages. “Promotions and Pricing Updates” ended the free models the Chinese list credited it with: for Qwen3.8-Flash, Hy3 and MiMo-V2.5 “The current free offer changes to a limited-time discount at 17:00 on September 16, 2026 (Singapore Time, UTC+8)”, billed from then “at 10% of the standard reference price”, and every model in “Pricing and Usage” carries a price. What remains is a referral gift, not a sign-up one: a friend who registers through an invitation link “can claim a dedicated invitation registration gift of 300,000 Credits” — $0.30 at 1,000,000 Credits to the dollar — valid 30 days. Login is by Web3 wallet or Google, and api.b.ai/v1/models answers 401 without a key (2026-09-18). Reopens if: A model priced at zero on the promotions page again, or a grant to anyone who signs up, without an invitation.2026-09-17 — A pay-as-you-go aggregator — 302.AI采用0月费和按用量付费模式，先充值，后扣费 (no monthly fee, pay by usage: top up first, then it is deducted) — and “The minimum recharge is $5 to use all AI products.” A $1 test credit exists on paper with changing conditions: a help FAQ last modified 2024-11-07 gives it to every user who binds a phone, and a changelog entry of 2025-03-03 narrows it to 新用户只有填写邀请码才可以获得1美金测试额度 — only new users who enter an invite code; the referral-code page that entry links to answers 404. Reopens if: A current 302.AI page grants new accounts API credit without an invite code or a top-up.2026-09-17 — The RouteLLM API comes with a subscription: “Sign up as a ChatLLM subscriber to access RouteLLM API”, and “A ChatLLM subscription is $10 per user per month. You will be billed $10 when you sign up.” No free API access is named. Reopens if: A RouteLLM API key is available to an unsubscribed account.2026-09-17 — A router sold in prepaid packs, with no account before payment: “Buy a credit pack at https://above.dev/pricing ($10, $20 or $50). Checkout creates your account, emails your API key, and opens the dashboard for usage and balance history.” Reopens if: An account or key obtainable before checkout, with a stated free credit.2026-09-17 — Genuinely free and legitimate — “A free, community-powered generation service: volunteers share spare computer power so anyone can generate images and text” — and OpenAI-shaped at oai.aihorde.net, where anyone may “use the anonymous API key 0000000000 at the lowest priority” (a call with no key at all answers 401). It cannot drive a coding agent: the route’s ChatCompletionRequest schema has no tools field, and what it serves is a live census of whichever volunteers are online, mostly small roleplay fine-tunes, with no model an agent can count on being there tomorrow. The route calls itself “a pilot and might be restricted or expanded in the future” (2026-09-17). Reopens if: AI Horde’s OpenAI route accepts tool definitions, or it publishes a set of models it keeps available.2026-09-17 — Open models hosted in Japan on prepaid credit: “ai& uses a prepaid credit model. You add credits to your organization, and inference requests deduct from that balance as they complete.” Redemption codes add credit without a card, but “Redemption codes are handed out at events, hackathons and workshops”. A cookbook page calls qwen/qwen3.6-27b “free for prototyping”, while the keyless catalog prices it at $0.32 in and $3.20 out per million tokens (2026-09-17). Reopens if: A zero-priced model in the keyless catalog, or a public signup credit.2026-09-17 — A governance gateway over the customer’s own provider credentials — “Connect a provider, create a project API key, and call the gateway.” — selling paid plans of its own, billed apart from inference: “This page describes inference cost attribution, not the price of an Aixy subscription.” It hosts no capacity to give away. The site answers 403 to scripts; its llms-full.txt is readable. Reopens if: Aixy-hosted model capacity with a free allowance that needs no provider key of your own.2026-09-17 — A European inference API hosted in Germany whose free user account gets trial credit with no card — “Business and professional users can create a free user account and receive trial credit for API evaluation. Depending on the models used, this covers millions of tokens. No credit card is required, and the trial does not automatically convert into a paid service.” — but no page gives the credit a figure or the trial a length. The terms say only that “The trial period ends as specified on the website when applying for the trial period”, and that the service “is reserved for entrepreneurs” acting in a trade, business or profession, with a company account (register number and VAT ID) needed once the trial ends. A credit with no figure gives a probe nothing that ends with it, the Baseten case. Reopens if: AKI.IO publishes the size or the length of the trial credit on a readable page.2026-09-17 — A verified-inference network whose API needs paid credit: “The Anthropic-compatible endpoint requires paid credits, like the OpenAI-compatible one: on the free tier requests return 402.” Its cheapest plan is $1 a month — “A small paid tier for trying Ambient.” — and the keyless catalog prices all four models. Reopens if: Ambient’s free tier or its drops can make API calls that do not return 402.2026-09-17 — A real daily quota behind an Alipay payment agreement. The pricing page: “Each account receives 500,000 free tokens per day (shared across input and output). The quota resets daily at 2:00 AM UTC+8 and does not roll over.”, on Ling-3.0-flash, Ling-2.6-1T, Ling-2.6-flash and Ring-2.6-1T at 2 QPS. But “Before your first API call, bind an Alipay account and activate the Ling service”, the FAQ adds “You need to sign the Mini Program Cloud auto-debit agreement and activate the Ling model service”, and “Only personal Alipay accounts are supported for payment.” An auto-debit agreement is a payment method on file, and whether an Alipay account opened outside mainland China can sign it is stated on no page read. Ling 3.0 Flash variants are already free on rows this list carries. models.dev still lists the service as Bailing at api.tbox.cn. Reopens if: The daily quota works without the auto-debit agreement, or a page says an Alipay account from outside mainland China can sign it.2026-09-17 — A router whose Free plan and “14-day free trial. No credit card required.” are platform tiers, with inference bought separately: “Pro is a flat $89/mo subscription billed separately from inference costs. You purchase credits at provider token prices with zero markup. Credits work the same on Free and Pro.” The keyless directory at api.auriko.ai/v1/directory/models prices five routes at 0, three to Z.ai and two to Google AI Studio, and no page says whether they answer an account with no credit. Reopens if: Auriko documents its zero-priced routes answering an account with no purchased credit.2026-09-17 — A Swedish inference API whose Trial plan has “€5 free starting credits”, but the pricing FAQ puts a card in front of them: “The Trial plan includes €5 in starting credits so you can try the API without a subscription — a card is required at signup to prevent abuse, but nothing is charged.” After the credit, “an active plan (from €25 / month in prepaid credits) is required for continued API access”. One-off credits behind a card are what CONTRIBUTING excludes. Reopens if: The €5 is granted without a card, or a free allowance recurs without a plan.2026-09-17 — Flat-rate subscriptions for coding agents. The account docs say “The first time you sign in, your account is created automatically on the Free plan, so you can try it before paying anything”, but no page gives the Free plan an allowance, and the API reference lists 402 “No active subscription”. The site’s source carries promotional strings for a signup credit and daily free credits that no rendered page shows. Reopens if: A page states the Free plan’s API allowance, or a signup or daily credit usable without a subscription.2026-09-17 — Sherlock, CloudFerro’s model API, is billed per token — “Our pricing model is based on token usage - both input (received) and output (generated) tokens for each model” — and “To use Sherlock’s API, you have to be an admin user of an Organization registered at CloudFerro”, an organization that needs an EU VAT number or tax ID. The cloud’s free trial (“We grant up to 250 EUR for testing purposes within our infrastructure”) is not said on any page read to reach Sherlock tokens. Reopens if: A CloudFerro page says the trial credits pay for Sherlock, without a card or a company tax ID.2026-09-17 — Gated by request: “Coral Inference is currently a design-partner program. To get access, contact hello@coralbricks.ai.”, and ungranted accounts get 403 access_denied. Usage is billed from a prepaid balance. Reopens if: Open self-serve access with a stated free credit or free model.2026-09-17 — “Usage-based pricing, one unified bill, no subscriptions.”, funded by Stripe top-ups with a 5.5% service fee — “There is no minimum fee — but the minimum top-up is $10.” No signup credit or zero-priced model is named. Reopens if: A signup credit or a zero-priced model on CrossModel’s own pages.2026-09-17 — The pricing page offers “$5 in free credits” to “Fine-tune and serve models in Crusoe Intelligence Foundry”, while managed inference is pay-as-you-go per million tokens, and the account docs put a card in front of both: “To provision resources or use the managed inference service, you must enable billing on your account. You will need a valid, non-prepaid credit card to proceed.” Reopens if: The $5 can be spent on inference before a card is added.2026-09-17 — A mainland platform whose token API is pay-as-you-go — MaaS by Token：使用 Token 计费，共享资源 — on an account that must pass real-name verification: 根据适用法律要求，您需要进行实名认证以使用我们的产品或服务。 The docs home still shows an undated “DeepSeek free for two weeks” banner with no terms behind it. Reopens if: A dated, current page grants free token usage or a signup voucher without mainland real-name verification.2026-09-17 — A new team gets a signup credit — “When you create your first team, DigitalOcean automatically applies a $5 signup credit” and “Signup credits expire 90 days after signup” — but a payment method comes first (“You must add a valid payment method before you can create Droplets or other resources”), and “Serverless inference is prepaid only. You must maintain a positive prepaid account balance to send serverless inference requests”. Whether the signup credit counts toward that balance is not stated. Reopens if: DigitalOcean documents the signup credit paying for serverless inference with no payment method on file.2026-09-17 — Open models hosted in Tokyo at per-token prices, with direct keys handed out on request and aimed at high-volume users; the site renders only in a browser, and no page names a free tier or credit (2026-09-17). Reopens if: A self-serve key with a stated free allowance.2026-09-17 — The model API spends purchased compute points — 在正式使用模型API服务前，需先购买算力点。算力点购买仅支持通过英博云账户现金余额进行支付 — bought before the first call and only from cash balance, on an account that has done 注册和实名认证, registration and real-name verification. The ¥20 the home page gives for registering and verifying (注册认证送20元无门槛代金券) is a voucher, where the points are bought with cash balance only. Reopens if: The model API accepts vouchers or offers a free allowance, without real-name verification.2026-09-17 — A workspace API for agent runs rather than a chat endpoint — “The workspace API has a saved-run contract; OpenAI Chat Completions SDK request fields do not apply.” — paid from workspace credit or with your own provider key: “No Tracer fee on the first $30,000/month in eligible usage per workspace with your own API key.” One model in its public library, the anonymous Union Alpha preview, is flagged free, with its availability unstated. Reopens if: A free credit for new workspaces, or a free model callable through the API without payment.2026-09-17 — A unified API whose keyless catalog prices 1,064 rows, a handful at 0 — google/gemma-4-31b-it, google/gemma-4-26b-a4b-it and cohere/command-a-plus-05-2026 among them — while billing is a credit system: “Eden AI uses a pay-per-use credit system”, credits are bought with a 5.5% platform fee, and the pricing page names no signup credit. Whether a zero-priced row answers an account with no credit is stated on no page read, the question NanoGPT’s record turns on. Reopens if: Eden AI’s docs say zero-priced models run on an empty balance, or a signup credit is named.2026-09-17 — evroc’s shared models are billed per token, and two posts from September 2026 offer a campaign: “For the first 100 signups, we are offering €300 credits to help you get a started building on evroc and using evroc Think.” The sign-up form’s own script saves a card before the account exists — “Save Payment Method (€0 charge)” — and the billing docs say “At least one card must remain on your account”. A first-100 grant behind a card is excluded twice over. Reopens if: evroc offers credits that are claimed without saving a card.2026-09-17 — API usage draws on a paid balance — “Your subscription payments and added funds become a USD balance.” — and in the plan table API access belongs to the $25 Pro plan; the Free plan’s 300 credits a month cover the vendor’s own agents and flows. Reopens if: Free-plan credits can be spent through an API key.2026-09-17 — Serverless models are paid from bought credit, their prices kept in the console (“Rather than mirror prices here (where they would go stale), the up-to-date rates live in the console”), and usage tiers follow purchases — “Please note that voucher redemptions do not count towards purchase”. No page read names a signup credit or a zero-priced model; the GMI Router’s free model selection routes to paid models. Reopens if: GMI publishes a signup credit or a zero-priced serverless model.2026-09-17 — The free Hobby plan’s “10,000 free requests” are observability logs, not inference; the gateway serves models from bought credits or your own provider keys, and a call with neither fails — the error guide’s case is “No Helicone credits, no BYOK”. Reopens if: Free inference credit for new accounts.2026-09-17 — “Only pay for what you run. Simple, transparent, and built for real usage. No hidden fees.” — six serverless models, every one priced in the keyless catalog, plus dedicated GPUs from $3 an hour with commitment discounts; no free tier, trial credit or zero price on any page read. Reopens if: A signup credit or a zero-priced model on an Inceptron page or catalog.2026-09-17 — A prepaid, per-token API behind a waitlist: “Access is prepaid. Credit is consumed per token at the rates shown at the time of each request. Requests are rejected when your balance reaches zero.” The terms also say “Inco is an experimental prototype.” No starting credit is named. Reopens if: Inco grants starting credit at sign-up.2026-09-17 — “Infer by Flow7 provides public paid access. Anyone can create an account; email verification and wallet funding are required before metered API use.” Accounts start at $0, every selector in the keyless catalog is priced, and “Serving suppliers remain private.” Reopens if: A free wallet credit, or a zero-priced selector in the public catalog.2026-09-17 — The plans page gives the default Standard plan “free, light daily access”, with pay-as-you-go beyond “the daily limit”, and the table on the same page describes Standard as “Continuous access with PAYG” with “No refreshes, pay only for what you use”. No figure for the free access appears on any page read. The keyless catalog prices all 35 rows and opens four to the lowest access tier — GLM-4.5-Air, Gemma 4 26B A4B, gpt-oss-20b and Llama 3.3 70B — the rest needing a higher one. Reopens if: io.net publishes the Standard plan’s daily free allowance with a figure.2026-09-17 — The offer exists only in the site’s JavaScript, and the same scripts disagree with it. The pricing page’s script says “Flexible token pricing, high usage limits, and postpaid billing—plus $1 in free credits to get you started!”, a partner campaign’s says “Sign in to get $1 in free credits instantly.”, and the home page’s quick start says “Top up your account with credits to enable API calls.” Every one of the 20 models in the keyless catalog is priced. The operator named in the recharge agreement is Magik Compute Technology (Shanghai) Co., Ltd. Reopens if: A server-rendered Jalapeno page states the $1 grant and that it is spent without a top-up.2026-09-17 — A Swedish integration company’s model API, currently DeepSeek V4 Flash behind the name Kloker: “Usage is billed per million tokens, excl. VAT. Input is €0.20. Output is €1.”, with keys by request — “Not on the platform yet? Email klokintegration@integrera.com and ask for a token.” The 30-day free trial on the home page belongs to its chat assistant, not the API. Reopens if: Klok grants a free token allowance for the AI-API.2026-09-17 — A Prague-hosted API whose one chat model is Qwen3.8 27B, beside Whisper and text-to-speech models, offering “€10 in free credits when you sign up to try our API for free”. Nothing on the home page or on any page its docs index lists says whether a card is needed before the credit is spent — the question that decides a one-off credit here, as it does on SiliconFlow. Reopens if: A Kosmik page says the €10 is spent without a payment method on file.2026-09-17 — A GLM-4.7 coding plan with a 30-day free trial described as a promotion — 用户在推广活动期间可申请30天免费试用 — at about 40 prompts per five hours, and 套餐仅限于指定的编程工具中使用 (usable only in the listed coding tools). The claim page is client-rendered and answers “please log in first” to a script, and the launch page’s own bundle limits the trial to 100 developers a day, so whether it can be claimed today, and with what account, cannot be read. Reopens if: A readable page shows the trial open now, with a sign-up a reader outside mainland China can complete.2026-09-17 — Four open models on idle enterprise GPUs, per token from prepaid credits: “You need a positive credit balance before you can make API requests.” The personal subscriptions ($10, $30 and $100 a month) include usage worth $20, $75 and $300, all paid; no starting credit is named. Reopens if: Lilac grants starting credit without a top-up.2026-09-17 — A French lab’s own models, AGI 01 Swift and Frontier, on a metered API — “Metered usage, no monthly minimum. Pay only for the tokens your app sends and receives.” Its $0 Free plan, “A restrained entry into AGI 01, LucidArc, and LucidCode”, covers the vendor’s own chat, desktop and CLI apps with unpublished limits (the CLI offers “Connect with /login: free, subscription, or API key.”), and nothing ties that plan to the API. Reopens if: LucidQuery says the Free plan’s allowance works through api.lucidquery.com with a key other clients can use.2026-09-17 — “Access 60+ open-weight models on a pay-as-you-go basis. No commitments, no minimums. Just top up credits and go.” The free Explorer plan in the pricing page’s data carries a daily energy allowance for the vendor’s studio but not API usage, and for plans like it the docs say “API calls are credits_only, even if you have energy left”. The coding plans that do include API usage are paid and “intended exclusively for businesses as defined by § 14 BGB”. Reopens if: Melious lets the Explorer plan or a signup grant pay for API calls.2026-09-17 — The vendor’s own repository, github.com/mlvoca/free-llm-api, now names the models and the terms: “a publicly hosted /api/generate endpoint based on the Ollama API” serving TinyLlama and DeepSeek R1 (1.5b), which “currently works without any kind of rate limit or API key”, and “Commercial use of this api is not allowed”. It names no operator beyond a Proton Mail address, and mlvoca.com still answers 403 to every fetch (2026-09-17). There is nothing here a probe can anchor on that dies with the offer — a README outlives the endpoint it describes — and the endpoint speaks Ollama’s generate route rather than chat completions, so the keyless check cannot call it either. Reopens if: mlvoca.com serves a readable page about the endpoint, or the service answers an OpenAI-compatible chat completions route.2026-09-17 — A gateway for coding agents billed from a balance: “Prepaid usage with no subscription. Add prepaid usage balance with a card.” Managed requests deduct the catalog provider cost and BYOK requests a 5% fee; no free allowance is named. Reopens if: Model Oracle grants balance at sign-up without a card.2026-09-17 — An OpenAI-compatible gateway whose free tier is a RapidAPI listing: modelishub.com/pricing says “Subscribe on RapidAPI (free tier, no card to start)”, and the listing’s BASIC plan is $0 with a hard limit of 100 requests a month. The same listing says tools and function calling are not supported on current plans and “Output is capped at 1024 tokens” — a lane that cannot call a tool cannot drive a coding agent, the AI Horde case. Reopens if: The free plan accepts tool definitions and allows more than 100 requests a month.2026-09-17 — A paid membership, not a free tier: “A closed community of builders sharing dedicated GPUs to run open models. Every member pays to be here”, paid by “Credit or debit card”, with a waitlist for spots. models.dev lists its seven models at cost 0 because the price is the flat subscription; the monthly token allowances per member are what it buys. Reopens if: NaN offers a membership or an allowance at no charge.2026-09-17 — Confidential inference on purchased credits (“purchase credits based on your needs”), every model in the keyless catalog priced. The one way in without a card is a stake: “Stake NEAR to fund private inference and deploy always-on IronClaw agents. Credits scale with your stake. No credit card required.” — credits paid by the staking rewards on tokens the reader must own. Reopens if: Starting credits for new accounts without a purchase or a stake.2026-09-17 — A router that distills small models from an agent’s own traces, not an open lane: “NeoSmith is currently free for design partners during early access”, and the self-serve route is a booked demo that ends with a pilot key — “Walk away with a 14-day pilot key (no credit card)”. A key issued after a sales call is not an offer a reader can take. Reopens if: NeoSmith issues keys with a free allowance through self-serve sign-up.2026-09-17 — Energy-priced inference whose trial asks for a card first: the quick start says “adding a payment method unlocks $1.00 in free trial credit (no charge required; the credit is valid for 30 days)”, and the pricing page “Add a payment method to get $1.00 in free credits — no charge required — then add more when you need them.” Reopens if: The trial credit is granted without adding a payment method.2026-09-17 — Company AI sold per person to businesses, with a price still unset — “We are not going to invent one to fill this page” — and no plan for one developer: “There is no single-seat plan and there will not be one.” Its “Free to try” has no published terms. Reopens if: Published trial terms that give an individual developer an API key with a free allowance.2026-09-17 — New AI-inference users get a one-time package — 注册有奖活动：现在注册 qiniu.com/ai 开发者账号，立即获得 300万免费 Token 资源包, three million tokens counted at DeepSeek-V3.1’s input price — issued after the first billed usage (使用服务并产生计费后自动发放). Two walls keep it off. The API FAQ says 目前AI产品仅支持中国大陆账号使用，暂不支持海外账号 — the AI products take mainland-China accounts only, not overseas ones — and a new user has a 24-hour grace period (24 小时免认证体验期) before real-name verification is required: the mainland real-name account this list does not carry, as for Baidu Qianfan and Volcengine. Reopens if: Qiniu opens AI inference to accounts outside mainland China without real-name verification.2026-09-17 — A reseller gateway paid by Alipay top-ups from ¥10, with email verification; the only zero-priced entry in its pricing data is an internal test model (启航测试模型，不计费，仅用于测试 — a test model, not billed, for testing only). No registration bonus is named. Reopens if: A registration bonus, or a zero-priced production model.2026-09-17 — A membership router: “Basic is $10/month and includes $10 in routing credit. Add more whenever you need it.”, and the API quick start requires “An active Basic membership with a positive Routing balance”. Every row in the keyless catalog is priced. Reopens if: A free tier or signup credit usable without the Basic membership.2026-09-17 — Hosted model APIs on a prepaid balance, with a one-off dollar: the docs say “Signing up is free and grants $1 once per account” and the pricing page “New accounts start with $1 free to test the Model APIs”, top-ups starting at $10 after that. A one-off credit is listed here when it asks for no card, and neither page, nor the pricing FAQ, says whether this one can be spent before a payment method is added (2026-09-17). Reopens if: A RunInfra page says the $1 is spent without a payment method on file.2026-09-17 — Paid subscriptions from $20 a month or pay-as-you-go credits — the terms read “To use the Service, You must purchase credits that may be used within the Service” — and “we do not provide services to users in EU (European Union) or EEA (European Economic Area) member states”. The console’s pricing page ships interface strings for a free usage allowance that no public page describes. Reopens if: A public Sakana page states a free API allowance for accounts without a subscription.2026-09-17 — A gateway in early access, paid per token from a credit balance or by a monthly subscription arranged with support: “The organization must have a positive credit balance before it can process AI Gateway requests.” The pricing FAQ answers its own question about a free trial: “No. SaladCloud and the Transcription API are pay-as-you-go with no minimum commitment”. Reopens if: Salad grants new organizations credit, without a card, that the AI Gateway accepts.2026-09-17 — China’s national supercomputing marketplace runs a promotion from 2026-08-13 to 2026-10-13 whose new-user package is 1000万 token 量包 — ten million tokens of GLM-5-Base, valid 30 days — under 仅限活动期间新注册用户，每个账号限领1次。福利总量有限，先到先得 (new accounts only, once each, while supplies last). A base model is not what a coding agent runs, new accounts are told apart by phone number, ID-card number and payment account, and the API docs say API access may have to be opened through a business advisor (如有需要，可联系您的专属业务顾问或平台客服开通调用 API 权限). The Token Plan itself is a paid subscription for AI tools only. Reopens if: A recurring or chat-model allowance that opens without mainland real-name verification.2026-09-17 — An Australian sovereign inference provider priced per token, with monthly plans in compute units from A$500; the site answers 403 to scripts and its rendered pricing and plan pages name no free tier, trial or credit (2026-09-17). Reopens if: A sign-up credit or trial for new accounts on an SCX page.2026-09-17 — A free 30-day test by application, for companies: “All registrations are manually checked by us. If you qualify, you will receive your personal token to use STACKIT AI Model Serving within 24 hours. The trial period ends automatically after 30 days.” The request form requires a company name, and a customer account needs to “Provide the company’s VAT ID and Commercial Register number.” Reopens if: Self-serve sign-up, without approval or a company, that grants a free token or credit.2026-09-17 — “Usage based pricing or monthly plans to power your agents efficiently.” — per-token credits from the dashboard or plans from $100 a month; no free plan or trial is named. Reopens if: Free starting credits or a free plan on the pricing page or in the docs.2026-09-17 — Per-token open models from a funded balance, and the FAQ rules out a trial in so many words: “We don’t currently offer refunds or trial credits due to the overhead of processing these requests.” Reopens if: Trial credits or zero-priced models in the InstaGen docs.2026-09-17 — EU inference on prepaid credits bought by card, with no free tier: “When your balance falls below $0.05, your account and API keys will be disabled.” Reopens if: A starting credit without payment, or a zero-priced model.2026-09-17 — Confidential inference whose API key comes after checkout: “Complete the checkout by entering your payment details. You’re only charged based on usage.” Every model in the keyless catalog is priced. Reopens if: A starting credit without payment details, or a zero-priced model.2026-09-17 — A fine-tuning API with sampling endpoints, billed by use — “Tinker uses a pricing plan that reflects usage. All prices are in USD per million tokens.” — whose serverless inference is a beta for Thinking Machines’ own Inkling models; research grants are the only credit mentioned. Reopens if: Free credits for new accounts without an application.2026-09-17 — “TokenGO uses a pay-as-you-go model. You only pay for the tokens you actually consume, priced per million tokens.”, funded through Stripe before the first request; the “Free” row in its rate-limit table is a speed tier, not a balance. Reopens if: A free balance at sign-up, or a public code that grants credit without payment.2026-09-17 — The keyless catalog prices trustedrouter/free, a routing pool, at 0 — beside orchestration aliases that also read 0 — but the pricing page bills every route: “Pay the provider price plus 5.5%, with no monthly plan”, and “Prepaid text and embedding prices are provider cost + 5.5%, with a $0.01 per million token floor”, paid from prepaid credits, BYOK or usage-based billing. Its only credits are migration credits for “Teams already spending more than $100/month on LLMs”, by approval. Reopens if: TrustedRouter exempts trustedrouter/free from the floor or publishes a free allowance.2026-09-17 — Open coding models on the vendor’s own GPUs, paid from a wallet: “A wallet with no credit at all — no paid top-up, no active promo grant — does not serve until it has some”, the first rate tier unlocks at “Your first top-up”, and promo credit such as the founders’ bonus serves only while it lasts. The keyless catalog prices one id at 0, umans-deepseek-v4.1-flash-lab, which the docs never name; the one lab they describe, DeepSeek V4 Pro, “served here as a seat-gated pre-release lab”. models.dev also lists an Umans AI Coding Plan at cost 0, and the docs speak of plans only as archived plans. Reopens if: Umans serves an account with no top-up, or documents a zero-priced model any account can call.2026-09-17 — The closest of the three to qualifying and still out. Its free plan is “$5 of included monthly credits” with a “7 message/day limit”, and it does have an API — api.v0.dev/v1, keyed by V0_API_KEY — but that API is /chats/:id/messages, a product endpoint for driving v0’s own builder, not a model endpoint. The credit buys v0 messages, not tokens you can spend on your own codebase. Vercel’s model credits that can be spent that way are already listed here as Vercel AI Gateway. Re-read 2026-09-17: a keyless POST to api.v0.dev/v1/chat/completions answers 404, while /v1/chats answers 401. Reopens if: The v0 free plan’s credits become spendable through an OpenAI-compatible or otherwise general model endpoint.2026-09-17 — Free credit for new accounts ended on 2026-09-02: “Effective immediately, newly registered users will no longer automatically receive complimentary API credits.” Since then the only credit is a one-time bonus after a first top-up of at least $5, and no model is priced at 0. Reopens if: The revised trial the 2026-09-02 announcement promises grants credit without a top-up.2026-09-17 — Three vision models paid in units bought in rupees — the app’s own text is “Buy units to continue — you’ll return here after payment.” — and priced in the keyless catalog; no free units are named. Reopens if: Free units at sign-up, or a zero-priced model.2026-09-17 — The Free plan is a platform tier rather than an allowance — “$0 to start Pay-as-you-go model usage” — and the credits on offer are the Growth plan’s “$25 in token credits included every month”, at $25 a month, and “$200/mo credits available to approved open-source maintainers on request”, a grant by approval rather than a tier anyone can use. The operator is Allegro US, LLC. Reopens if: The Free plan includes model usage with a figure.2026-09-17 — A paid subscription: the docs provision “a Vultr Serverless Inference subscription” before a key exists, and neither they nor the API reference at api.vultrinference.com name a free allowance. vultr.com’s pricing pages answer 403 to a script (2026-09-17); the keyless catalog lists 16 models. Reopens if: Vultr publishes a free inference allowance that needs no subscription.2026-09-17 — “Access Wafer models through an OpenAI-compatible API. Add credits, create an API key, and pay per token.”, or a paid Wafer Pass subscription. Its $500 of free credits belong to a startup program that begins with “a 25-minute conversation with our team”. Reopens if: Credits for any new account without an application, or a free Wafer Pass tier.2026-09-17 — Paid entry only: “Choose monthly capacity from $20—or try the API with $2 prepaid credits.” Every priced model in the catalog is paid. Reopens if: Credits granted without payment.2026-09-17 — A Brazilian platform whose sign-up credit reaches its AI API with no card — “Ao criar sua conta, você ganha R$30,00 em créditos automaticamente” and “Você não precisa ter cartão de crédito para começar” (a new account gets R$30 of credit automatically; no credit card is needed to start), usable on AI (“Os R$30,00 recebidos ao criar a conta também podem ser usados com a oferta de IA”) — but registration asks for a “CPF ou CNPJ válido” and “residência no Brasil”, a Brazilian tax ID and residence: a national identity wall, like the mainland real-name accounts recorded above. Reopens if: Zenifra opens registration without a Brazilian tax ID and residence.2026-09-16 — 01.AI closed its model API: api.lingyiwanwu.com answers GET /v1/models with HTTP 410 {“error”:{“code”:”model_service_closed”,”message”:”Model service has been discontinued”}}. The company is alive on the same domain — www.lingyiwanwu.com announces “2026.07 万策平台发布”, an enterprise decision platform — so the offer ended, not the service. It sat on the blocklist from 2026-09-14 as a product that was gone, which would have buried a vendor that can reopen an API. OmniRoute’s feed still names its “Yi-Light” free models. Reopens if: api.lingyiwanwu.com/v1/models serves models again, or 01.AI publishes API pricing with a free allowance.2026-09-16 — A paid relay with nothing free found and no cause shown. It sat on the blocklist from 2026-09-14 on its own discount claim (“85–90% off Claude, 90% off GPT”), and a price is not provenance. Its FAQ separates “claude-“ models, discounted and “best suited for coding scenarios”, from “anthropic/” models “Routed through AWS”, without saying where the first kind come from. aiberm.org is its mainland mirror. Reopens if: Aiberm offers a free allowance, or states where its discounted Claude capacity comes from.2026-09-16 — “The Free Tier is currently paused” (docs.aimlapi.com/faq/free-tier, 2026-09-16), though the marketing pages still offer “Free credits on signup”. Reopens if: The free tier resumes, with its quota stated.2026-09-16 — Its own FAQ answers the question: “There is no free tier for Rovo Dev”, and “Rovo Dev Standard is available with a free 30-day trial” (atlassian.com/software/rovo-dev/pricing, 2026-09-16). The limited monthly credits Atlassian extends to Jira Cloud customers ride on a paid subscription, and neither page says whether the trial takes a card. Reopens if: Rovo Dev gains an allowance reachable without a paid plan, or a trial page states it needs no card.2026-09-16 — A one-off trial, readable now in the docs’ page data: ModelArk has a mode in which “calls to the inference API consume only the 500k free tokens granted by the platform”, and the campaign terms grant “500,000 (Five Hundred Thousand) tokens per Model”, “redeemable only once per BytePlus account”, after “BytePlus account registration and enterprise information submission”. Sign-up has you “select one of the options as your payment method”, and no page read says the tokens can be spent before one is added (2026-09-16). Reopens if: A page says the free tokens need no payment method, or the grant recurs.2026-09-16 — Listed from 2026-07-19 to 2026-09-16 as a trial marked with a card, and taken off because CONTRIBUTING excludes one-off credits that require one. inference-docs.cerebras.ai/support/rate-limits: “New accounts receive $5 in free credits after adding a verified payment method. These credits expire 30 days after they’re granted”, “If you skip adding a payment method at sign-up, Playground and API access remain inactive until you do”, and “Is there a permanently free tier? No.” Reopens if: Cerebras grants credits without a payment method, or a free allowance that recurs.2026-09-16 — cto.new has a coding surface now: its changelog reads “Launched the first version of the CTO CLI” (June 5th 2026), and the free tier stands in the fair-use doc: “Use cto completely free. Access to the best value models with generous usage limits that reset every day”. No page read says whether the CLI runs on that free tier, and cto.new itself answers ordinary clients with a Vercel Security Checkpoint, so docs.cto.new is the readable surface. It is Engine Labs’ product (every docs asset is served from mintcdn.com/enginelabs/). Reopens if: A cto.new page says the CLI works on the free tier.2026-09-16 — iFlow stopped its developer service on 2026-04-17. Its own forum announcement that day reads “今日11点 iFlow CLI 将停止各项服务，除仅支持使用自定义模型外相关的模型调用、搜索服务都不可用” (from 11:00 today iFlow CLI stops its services; model calls and search are unavailable, only custom models remain). This record used to call a rival list’s report of that shutdown wrong; the report was right. Reopens if: iFlow reopens model calls with a free quota.2026-09-16 — Listed from 2026-09-02 to 2026-09-16 as a trial marked with a card, and taken off because CONTRIBUTING excludes one-off credits that require one: “One million free credits allow you to test the service without commitment for one month” and “A credit card is required to start using the API” (support guide 2845). Reopens if: The trial credits stop requiring a card, or a free allowance recurs.2026-09-16 — A free tier of roughly 200 requests a month on the fast-apply model is reported by third parties, and nothing readable confirms it: morphllm.com/pricing answers HTTP 429 with a Vercel Security Checkpoint, and docs.morphllm.com/llms.txt carries no prices and no free tier (2026-09-16). Reopens if: Morph states its free quota on a page an ordinary client can read.2026-09-16 — Paid from the first call: docs.nscale.com, readable again on 2026-09-16, opens with “Add credit to your account From the dashboard, add a minimum of $5 of credit to start using our service”, and no page names a free grant. Reopens if: Nscale publishes free credits or a free serverless allowance.2026-09-16 — Listed from 2026-07-20 to 2026-09-16 as a free tier marked with a card, and taken off because CONTRIBUTING excludes one-off credits that require one. “Every new customer gets 1,000,000 free tokens—start paying only from the 1,000,001st token” is a once-per-customer grant, and Scaleway’s account docs say “Ordering Scaleway resources requires a valid credit card”. Reopens if: The free tokens stop requiring a card, or become a recurring allowance.2026-09-16 — Being wound down. “Sunsetting Supermaven” (supermaven.com/blog, Nov 21, 2025): “We will provide free autocomplete inference for existing customers for the foreseeable future.” The pricing page still shows “Free Tier $0 /month”, but the free promise covers existing customers, and it was completions only — no chat, agent use or API. Reopens if: Supermaven takes new users again with a free plan that includes chat, agent use or an API.2026-09-16 — No free plan to read any more: “Tabnine has been acquired by Tricentis” (tabnine.com/pricing), and the subscription docs list only “Enterprise (SaaS)” and “Enterprise (private installation)” (2026-09-16). Reopens if: A free or individual plan returns.2026-09-16 — api.xkiro.com/v1/models now marks 42 rows access_tier free at 0/0, no longer only Qwen: openai/gpt-5.3-codex-spark, mistralai/mistral-large-2512 and mistralai/codestral-2508 are among them, and the site’s own JSON-LD featureList includes “Claude and GPT proxy”. Free closed models prove nothing about where the capacity comes from, and no page states it either — no upstream provider is named anywhere read — so it stays unlisted until that is known (2026-09-16). Reopens if: xKiro names where its free models are served from, or a provider confirms the arrangement.2026-09-14 — Moved here from the blocklist, where it sat as “BYOK-only, no bundled free model usage”. That is true of aider and it is still a verdict on an offer, not on the tool — Cline sat on the blocklist under the same words for two months while its own provider handed out free models, because nothing on that list ever comes up for a re-check. aider ships no provider of its own: its model docs send readers to other vendors’ free tiers — “Aider works with a number of free API providers: OpenRouter offers free access to many models, with limitations on daily usage”, Gemini beside it — which this list carries directly. The repository was last pushed on 2026-05-22. Reopens if: aider ships a hosted provider or a sign-in of its own that includes model usage.2026-09-14 — Free today only with inference you bring. The Hobby tier announced on 2026-09-13 is “Free” with “Use your ChatGPT sub & other subs”, “Bring your own keys (BYOK)” and “Pay-as-you-go for orbs”, and the announcement says it outright: “Amp is now free to use when you bring your own compute and model subscriptions/keys … You can still pay for model inference through Amp if you want, with no markup.” The bundled lane Amp was known for is gone: its 2025-10-15 Amp Free post now opens “Update 4: Amp Free is paused. We don’t intend to bring it back”, after the ad-funded $10 a day of 2026-01-08 and “Amp Free Is Full (For Now)” on 2026-02-10. Reopens if: Amp brings back a free credit grant or a free mode that needs no subscription or key of your own.2026-09-14 — A mainland console with nothing readable: platform.baichuan-ai.com renders 108 characters, and api.baichuan-ai.com/v1/models answers api_key_empty. OmniRoute’s own correction is that “current reality is only a one-time 80 CNY trial credit for new users (valid 3 months)” — a trial, and on the same mainland-registration terms that keep the other Chinese consoles off this list. Reopens if: A readable page states the grant and a registration open outside the mainland.2026-09-14 — A coding agent for open models whose cheapest plan is paid. The home page FAQ says “Free tier for solo developers”, but the pricing page starts at “Go $1 /month + processing fee” with $10 of credits, then GOAT at $10/mo with $70. The free models it advertises — “Free on Laguna S 2.1 Free on Ling 3.0 Flash Free on LongCat 2.0 Free on Ling 3.0 Flash Sante”, “Requests cost no credits while capacity lasts” — are listed inside those plans, not beside them. Reopens if: Command Code offers a $0 plan with model usage included.2026-09-14 — An EU-compliance routing gateway whose Free plan prices the routing, not the tokens: “Free €0 /mo Pay for what you use Monthly limit: 10K requests/month Rate limit: 60 RPM 15% markup”, and its keyless catalog prices every one of its 147 rows. Reopens if: A model in the eurouter.ai catalog is priced at zero, or the Free plan includes token usage.2026-09-14 — No free plan. Individual plans start at “Pro $20 / mo”, then Plus at $100 and Max at $200, and the free-sounding parts are inside them: “Droid Core: free open-weight model pool with separate Rate Limits after Standard Usage runs out” and “BYOK is free up to an allowance on all Individual plans” both describe what a subscription includes (docs.factory.ai/pricing/individuals). Reopens if: Factory offers a $0 plan whose Droid Core or other model usage needs no subscription.2026-09-14 — Operated by AI.tech Ltd under New York law, with a real :free lane in a keyless catalog — openai/gpt-oss-120b:free, openai/gpt-oss-20b:free, google/gemma4-26b:free, nvidia/nemotron-3-nano-30b:free, nvidia/nemotron-3-super:free and sarvam/sarvam-105b:free, all priced 0 — that a new account cannot use. Its own docs set the gate: “up to 10 requests per org per day, with no payment required for organizations having paid credit balance >$1”, and “Free models cannot be used when your organization’s paid credit balance is <= $1”. The signup grant the site advertises — “Get Millions Of Tokens To Start. Free Credits On Us” — carries no figure, and the $0 Starter plan is a gateway plan for your own keys (“1 BYOK key · 1M requests / month”). Reopens if: The :free models answer an account that has never paid, or the signup credit is published as a figure.2026-09-14 — Subscriptions only: “Chat For roleplay, stories and long conversations … $25 /month” and “Developer For coding agents and API traffic … $50 /credits per month”, over a catalog of 40,000+ open models every row of which is priced. OmniRoute lists it with a free tier and then corrects itself — the only free access is an application-gated builder programme. Reopens if: Featherless publishes a free plan or a free model lane.2026-09-14 — A consumer suite of 477+ tools whose free allowance is real and small in the way that matters here: “Anonymous users get 6,000 tokens/day. Sign up free for 30,000 tokens/day”, spent on “Qwen 2.5 for chat” — 2024-era 7B to 72B models on its own GPUs — while “GPT-6, Claude, Veo + 584 premium models” start at $5. The API sits at api.free.ai/v1/chat/ with ids like qwen7b rather than a model a coding agent would be pointed at. Reopens if: The free daily allowance covers a current coding model on the API.2026-09-14 — A crypto-native agent platform — “Sponsored tips funded with USDC”, spending permissions “on a Base wallet” — whose opengateway catalog prices 13 of its 14 ids and zeroes one, nvidia/nemotron-3-ultra-550b-a55b:free, a model this list already carries free on five rows. OmniRoute’s note is that its free MiMo access “was removed in May 2026” and the Nemotron row is “a temporary promotional model”. Reopens if: gitlawb publishes a free lane of its own beyond one promotional id.2026-09-14 — EU-hosted chat on renewable energy, where the free part is a trial and the API is paid: “All API usage requires an active Pro, Teams, or API-only subscription”, every plan comes “with a 14-day free trial included”, and the entry plan is €4,50/month for 15 prompts a day. Reopens if: GreenPT’s API gets a free allowance outside the subscription trial.2026-09-14 — CAST AI’s open-source coding agent. The $0 Community tier is the harness alone — “Bring your own inference - Kimchi’s managed API, an OpenAI-compatible endpoint, or any LLM provider. No hosted inference” — and model usage starts on Coder at $20/mo, which “Includes $25 monthly credits”. The “no credit card” on the home page is about installing the CLI. Reopens if: The Community tier includes hosted inference or credits.2026-09-14 — “Free LLM Inference. No Limits”, with the price stated plainly — “In exchange for free inference, we log every request — prompts, completions, and metadata. After PII scrubbing this may be used in our private, internal evaluation datasets” — and an operator described only as “an independent research and engineering group”; /terms and /about answer 404, and the privacy policy names no company. The keyless catalog is 19 ids, and 14 of them — glm-5.3, kimi-k2.6, qwen-3.8-27b, gemma-4-26b, melotts, aura-2-en, nova-3, lucid-origin, phoenix-1.0 and the flux and whisper builds — are Cloudflare Workers AI’s names, whose free allocation is 10,000 neurons a day per account. Unlimited free use of it is either paid for by someone unnamed or pooled. Reopens if: Logfare publishes terms naming its operator and a stated quota, or says how the inference is paid for.2026-09-14 — Unreachable: monsterapi.ai and api.monsterapi.ai both answer SERVFAIL on Google’s public resolver (2026-09-14), so no page or catalog can be read. The free tier OmniRoute last described was one-time trial credits with “0 credits/month” recurring. Reopens if: The domain resolves again and a readable page states a free allowance.2026-09-14 — The free plan is real now and published, and what it serves is the reason it stays here. The keyless router.bynara.id/api/plans reads “Free tier with a daily token quota. No card required”, token_cap_daily 7,000,000 at 15 requests a minute, over four models — agnes-2.5-flash, laguna-s-2.1, stepfun-3.7-flash and tencent-hy3-free — which are other vendors’ free lanes: Agnes prices its Flash at zero itself, Laguna S 2.1 is free on OpenRouter and Kilo, and Step 3.7 Flash on Kilo. The paid day passes beside it (Freemium at Rp 5,000 a day, FreeMium Max, FreeMium Ultra) sell glm-5.3-free, mimo-v2.5-free, qwen3.8-flash-free, deepseek-v4.1-flash-free and muse-spark-1.3-contributor-free — ids named for free lanes, sold by the day. OmniRoute also reports that a key requires linking Telegram. The 2026-08-30 reading here found no pricing and a 401 catalog; this one was prompted by OmniRoute citing the plans endpoint. Reopens if: Nara’s free plan serves a model it hosts or licenses itself, rather than another gateway’s free lane.2026-09-14 — OmniRoute’s changelog says the free tier “is a recurring monthly free plan (10,000 requests/month)”, and nothing on the vendor’s side can be read to confirm it: nlpcloud.com/pricing answers a Cloudflare “Just a moment… Enable JavaScript and cookies to continue” challenge from this network and from a GitHub runner alike (2026-09-14), and api.nlpcloud.io/v1/models is a 404. The same OmniRoute file marks its terms “avoid” for proxy use. Reopens if: A pricing or docs page an ordinary HTTP client can read states the free plan.2026-09-14 — OpenAdapter, Inc. sells a coding subscription — “4x Cheaper than Claude Code”, Lite at $9/mo for “~400 requests over 5h” — across 72+ open models, and does have a free plan: “Try OpenAdapter Strict daily quota across the full catalog. No credit card.” The quota is never given a number, and the pitch is resale on its own words: “Same models. Fraction of the price. One subscription replaces 3+ separate provider plans”, beside an “0G Network” section of privately hosted models. Nothing says which provider plans are being replaced, or under what licence. Reopens if: OpenAdapter publishes the free plan’s daily quota and where its models are served from.2026-09-14 — The free hosted tier bundles no model usage: “Individual Free — Bring your own key or use our providers at-cost”, ten conversations a day, and the FAQ spells out the second half — “the OpenHands LLM provider (available within the Individual Tier), which provides multiple model options at cost, with no markup on a pay-as-you-go basis”. The agent is hosted free; the inference is yours to pay for or bring, the aider case. Reopens if: OpenHands Cloud includes model credits or a free model on its Individual tier.2026-09-14 — Moved here from the blocklist, where it sat as a browser-only SDK with no HTTP endpoint a coding agent could use. That stopped being true: Puter now serves an OpenAI-compatible endpoint at api.puter.com/puterai/openai/v1 and an Anthropic-compatible one at api.puter.com/puterai/anthropic, with its own tutorials for Claude Code, Cline and Kilo Code, and “every Puter account includes an allowance of storage, database, and AI usage at no charge”. The two do not meet. Its rate-limits page says the compatible endpoints “additionally require a paid plan — a free account calling them gets 402 subscription_required”, while the free allowance runs through puter.ai.* and /drivers/call, which no coding agent speaks, and the size of that allowance is shown only in the dashboard. The site’s “Free, Unlimited Claude API” tutorials are the marketing half of the same site, not what its docs describe. Reopens if: A free Puter account can call the OpenAI- or Anthropic-compatible endpoint without a paid plan.2026-09-14 — The sign-in that made Qwen Code a free agent is gone, in its own docs: “The Qwen OAuth free tier was discontinued on 2026-04-15. Existing cached tokens may continue working briefly, but new requests will be rejected.” The /auth menu now offers Alibaba Cloud’s Coding Plan (“weekly quota included”, for a fixed monthly fee), a Token Plan, a Model Studio API key or a third-party key, which leaves a BYOK agent. Alibaba’s free Model Studio quota is carried here as a row of its own. Reopens if: Qwen Code’s auth menu offers a sign-in with a free quota again.2026-09-14 — Moved here from the blocklist for the reason aider was: BYOK-only describes what is offered today, not the company. Roomote is the cloud coding agent the Roo Code team shut Roo Code down to build, and it bundles no inference — “Every plan includes app hosting and a sandbox provider — you just bring an inference key from any provider we support”. What is free is the hosting: “a free 7-day trial, no credit card” on Roomote Cloud, then from $49/mo, and self-hosting that “is free for up to 10 users”. Neither buys a token. Reopens if: A Roomote plan, trial or self-hosted tier includes model usage instead of asking for an inference key.2026-09-14 — A $1 monthly allowance with no card — “Free $0.00 /mo … $1 of model usage / month 10 requests / minute 1 API key Access to all open models” — on a gateway that markets “27 frontier models” and serves seven from its keyless catalog: nemotron-3-ultra, nemotron-3-super, gpt-oss-20b, laguna-xs-2.1, llama-3.2-11b-vision, flux.1-dev and an embedder, ids spelled the way NVIDIA’s API catalog spells them. Prices sit above the makers’ (DeepSeek V4 Flash at $0.27/$1.10), which is not the relay pattern, but the terms (June 2026) name no company and nothing says whose capacity the allowance spends. Reopens if: Speka names its operator and where its models are served, and the catalog carries what the home page claims.2026-09-14 — Every model on platform.stepfun.com/docs/zh/guides/pricing/details is priced, and the page records its free lanes ending rather than starting: “Step Audio R1.1 已升级为 Step Audio R1.5。该模型已结束限时免费” (the limited-time free period is over) and “step-2x-large 已于 2026 年 06 月 12 日 结束限时免费” (ended on 2026-06-12). The home page sells the Step Plan coding subscription. Step 3.7 Flash is free through Kilo, which this list carries. Reopens if: StepFun’s pricing page shows a zero-priced model or a signup grant.2026-09-14 — A one-off trial on Sweep’s own models with nothing readable about its terms. sweep.dev/pricing offers “Free Trial Free Try Sweep risk-free 1,000 Autocompletes $5 API credits included”, then Basic at $10 a month for unlimited autocomplete, Pro at $20 and Ultra at $60. The documentation it links answers HTTP 402 (docs.sweep.dev, 2026-09-14) and its terms and blog URLs serve the home page, so neither the trial’s conditions nor whether it takes a card can be read. Reopens if: Sweep publishes the trial’s terms on a page that answers, card or no card, or adds a recurring free allowance.2026-09-14 — A flat subscription for open models — “Subscription Packs … $1 /day $30 /mo Rate Limit 500 requests/5hr” — with the alternative “All models are pay-per-token”. No free plan or grant on the pricing page. Reopens if: Synthetic offers a free plan or signup allowance.2026-09-14 — The community fork that carried on when Roo Code shut down, and the one Roo Code’s own shutdown notice recommends. Its release notes invite users to take the Zoo Gateway key “to any OpenAI-compatible client or workflow”, but the gateway is prepaid — the terms read “Zoo Code credits are prepaid units” and “Minimum purchase amount is $5 USD”, non-refundable, and name no company. The public catalog behind zoocode.dev/models serves 251 models at Vercel AI Gateway’s own rates plus a ninth (minimax/minimax-m3 at $0.3333/$1.3333 against Vercel’s $0.30/$1.20), and the seven it prices at zero are exactly Vercel’s seven — the Ling 3.0 Flash Fin, Sante and VL ids and laguna-s-2.1-free — already on this list under Vercel AI Gateway. The v3.82.0 release notes promise “free access to MiniMax-M3 through Zoo Gateway” for a limited time while the same catalog prices it, and no page says an account without credits can call anything. Read 2026-09-14. Reopens if: Zoo’s docs or terms state that an account without purchased credits can call Zoo Gateway models, or MiniMax-M3 appears at zero in its public catalog.2026-09-14 — The Basic plan is real — “$0 /mo 7.2k req/day · 10/min 200k Daily Tokens Basic models only … No credits”, no card — and it can only reach the rows priced at zero, since Basic carries no credits. Those rows, in the keyless catalog’s BASIC tier, are nemotron-3-ultra, mistral-large-3 (675B), qwen-3.5 (397B), qwen-3.5-flash, mistral-small-4, gpt-oss-20b, step-3.7, diffusiongemma and llama-3.2-3b, plus Zylo’s own routers — the models NVIDIA’s API catalog serves free for development, handed out here under a gateway’s free plan without a word on where they run. Terms on zyloai.net/terms cover only linked-account enforcement. Reopens if: Zylo states where its zero-priced Basic models are served from, or the Basic plan covers a model it serves itself.2026-09-12 — The one text-LLM row of pacocartones/free-llm-api-hub left without a verdict here, because nothing under the domain answers. clarifai.com and www.clarifai.com time out on 443 and on 80 alike — from this network on 2026-09-05, 2026-09-11 and 2026-09-12, and from a GitHub runner on 2026-09-12, which is what takes a local block off the table — while api.clarifai.com, docs.clarifai.com, status.clarifai.com and portal.clarifai.com every one answer NXDOMAIN on Google’s public resolver. The domain’s nameservers are dns1 and dns2 at registrar-servers.com, the registrar’s own default pair, and www resolves into static.rcn.com. A platform whose API subdomain does not exist in DNS cannot be verified to offer anything, which is the rule that settled Coze and Morph LLM. It is here and not on the blocklist because none of this is a verdict on the company. Reopens if: clarifai.com or api.clarifai.com answers an ordinary HTTP client again, and the page it serves states the free tier.2026-09-11 — Named by cuihuan/awesome-ai-gateway’s relay watch-list as carrying “a free GLM/Gemma tier”. allrouter.ai and /pricing are a 2 KB client-rendered shell, /docs is readable (23 KB) and holds no “free” wording at all, and api.allrouter.ai does not resolve (2026-09-11). Unreadable rather than absent. Reopens if: A page this repository can read states the free tier.2026-09-11 — “An inference API built by Fastino Labs”, surfaced by the provider table of mnfst/llm-gateway. Every model on the home page is priced per token — Claude Sonnet 5 at $2.00 / $10.00, GPT-5.5 at $5.00 / $30.00, Gemma 4 12B at $0.25 / $0.25, GLiNER2 Large at $0.20 / $0.20 per 1M — and pioneer.ai/pricing sells seats: “PRO $20/seat/month, $40/seat/month in platform credits included”, Enterprise $50/seat. No signup credit and no free lane on either page as rendered here on 2026-09-11; the only “trial” in the page text is the name of its web font. Reopens if: pioneer.ai/pricing states a free tier or a signup credit.2026-09-11 — The product behind ClawLabsAI/free-ai-models, whose README funnels readers to “one OpenAI-compatible endpoint that auto-routes every request to whichever free model is answering best”. zerolimitai.com/developers does hand out a key — “OpenAI-compatible · Free tier”, base URL https://www.zerolimitai.com/api/v1, model “auto” — but the free key “expires after 7 days” and carries no stated quota; the paid tiers are “Lifetime Core $49 one-time … 2,000 API calls/day” and “Lifetime Pro $99 … 10,000 API calls/day”, sold beside “Unlimited chat”. What the key routes to is “the best free AI model available — Llama 4 Scout, DeepSeek R1, Qwen3 235B”, the models other vendors serve on their own free lanes, and neither /terms nor /about says where those requests go or under what licence; /docs answers 404. A relay reselling lifetime access to upstream free lanes is the provenance question llm7.io was blocked over, but the models here are open-weight and the operator publishes terms and a named team, so it stays a verdict rather than a block. Read 2026-09-11. Reopens if: It publishes the free key’s quota and where its “free models” are served from.2026-09-05 — A separate surface from the archived AI21 Labs (Jamba) entry. The 410 that retired AI21 Studio’s catalog route points here — “The AI21 Gateway is available at https://app.ai21.com” — and what is here is a bring-your-own-key gateway, not a model API: the app’s own env file names it ai21-intelligent-gateway-webapp with a base of api.ai21.com/gateway, its code snippets pass “your own OpenAI key” or “your own Anthropic key” upstream, and the only “Free trial” in it is a Stigg-billed “Tokenwise” plan trial counted in days. api.ai21.com/gateway/v1/chat/completions answers 400 “bad or missing authentication” keyless, so the route exists; no page, docs entry or llms.txt line describes it, and the docs index still lists only the Jamba, Maestro and Studio pages. Studio itself — where the $10 no-card trial lived — redirects every route to the www.ai21.com homepage. Reopens if: AI21 documents the Gateway with a model AI21 hosts itself and a free quota or credit on it that needs no card.2026-09-05 — The China-site twin of the listed Alibaba Cloud Model Studio row, reached through the models.dev digest as alibaba-cn. Its new-user quota is documented in full at help.aliyun.com/zh/model-studio/new-free-quota, which also serves as .md: “每个模型均有独立的免费额度（通常为 100 万 Token）” — an independent quota per model, usually one million tokens — for “90 天” from activation, in the 华北2（北京）region alone, and “无需实名认证即可获取和使用免费额度”: no real-name verification is needed to receive and spend it, an unverified account being stopped with a 403 AllocationQuota.FreeTierOnly when it runs out. Two things keep it off the list. The first-API-call guide on the same site says “如果开通服务时提示’您尚未进行实名认证’，请先进行实名认证” — activation may ask for the verification the quota page says is not needed — and the account sign-up and activation pages are client-rendered, so which of the two a reader outside mainland China meets cannot be read. The international edition, with its own quota, is the row this list carries. Reopens if: A readable vendor page says an aliyun.com account can be opened with a non-mainland phone number and Model Studio activated without real-name verification — the quota itself is documented and large.2026-09-05 — A Chinese list credits it with a daily $10-equivalent quota on AMD-hosted open models, OpenAI-compatible, no card, run with ZZ.ai and limited-time. The page at /radeon/tokenfactory is client-rendered: the readable part (2026-09-05) is a help notice for the “AMD AI 开发者计划”, the Developer Cloud and a joint ModelScope incentive program, with no quota, model, endpoint or sign-up term in it. The domain is AMD’s China developer site. Reopens if: A server-rendered page with the quota and the sign-up terms.2026-09-05 — claude.com/pricing (read 2026-09-05) prices the API per token from “$1 / MTok” on Haiku 4.5 to “$10 / MTok” on Fable 5.1 and names no free tier, trial credit or free model; the only free line is operational — code execution has “50 free hours of usage daily per organization” — and “Save 50% with batch processing” is a discount. Two of the lists read still link console.anthropic.com as a source of trial credits; the page no longer carries one. Reopens if: A free tier, free model or signup credit on the pricing page.2026-09-05 — The pricing page (read 2026-09-05) still sells a Lite plan “Free forever” — “20 req/minute”, “200 req/day” on small models, “10 req/day” on large ones, “Unlimited Tokens!” — over a catalog that stopped in 2024 (Llama 3 and 3.1, 8B and 70B). api.awanllm.com timed out on every call that day, catalog and chat alike, at 40 seconds, so the offer cannot be shown to answer, and a plan whose API does not respond is not one this list can verify twice a week. Reopens if: api.awanllm.com answers a request.2026-09-05 — The new-user quota is documented at cloud.baidu.com/doc/qianfan/s/Imi2rpirg (updated 2025-11-17, read 2026-09-05): from 2025-10-24, a user who has “阅读并同意用户协议” is granted “100万” tokens per model with “3个月” validity on 17 models — ERNIE-4.5-Turbo and ERNIE-X1-Turbo, DeepSeek-R1 and V3.1, Kimi-K2-Instruct, Qwen3-Coder-480B-A35B-Instruct among them — for online inference only. A one-off three-month grant is a trial rather than a standing lane, and the page says nothing about who may open the account; Baidu AI Cloud’s sign-up is the mainland real-name flow this list cannot read, beside Alibaba’s China site for the same reason. Reopens if: A readable page saying the console opens without mainland real-name verification, or an international console carrying the grant.2026-09-05 — Reached through the models.dev digest (2/2 rows at cost 0, nemotron-3-ultra-550b-a55b:free and gemma-4-31b-it:free). bothub.ru/models lists nine :free ids with no price where every other row carries rubles per 1M tokens, and its FAQ says where they are free: “Есть бесплатные модели с постфиксом ‘:free’ и ‘-exp’, которые можно использовать бесплатно через мини-окно на главной странице, а также странице модели” — through the mini-window on the home page and the model page, which is the web chat. The API documentation at bothub.ru/api/documentation/ru says “вы оплачиваете только фактическое использование API и моделей” and names no free tier, the model pages count a CAPS balance, and the catalog at openai.bothub.ru/v1/models answers 401 keyless. Reopens if: BotHub’s API docs say the :free ids cost 0 CAPS through the API, or the catalog opens keyless with those rows priced 0.2026-09-05 — Reached through the models.dev digest (2/109 rows at cost 0). The keyless catalog at api.cortecs.ai/v1/models prices every row in EUR and the only two at 0.0 are qwen3guard-gen-0.6b and qwen3guard-gen-8b, safety-moderation classifiers served through OVH — the same shape as OVHcloud’s zero-priced guard rows, and not a chat model. cortecs.ai/pricing is a prepaid balance with a “5% flat service fee”, paid “via credit card or by invoice” behind an Auto Top-up threshold; no signup credit is named anywhere on it. Reopens if: A chat model priced 0 in api.cortecs.ai/v1/models, or a signup credit with a figure on cortecs.ai/pricing.2026-09-05 — “A free community endpoint” for GLM 5.3 Flash from an “Independent AI research lab” that says it is “built in Germany”, with weights on Hugging Face under empero-ai. On 2026-09-05 free.empero.org is a maintenance page — “We are switching the free endpoint to new models” and “We are preparing the free endpoint for Qwen3.8-Flash-Next” — and no terms, rate limit, sign-up or catalog page could be read. Nothing here says how it is hosted or funded. Reopens if: The endpoint answers and a page states its limits and terms.2026-09-05 — The API platform behind the Kimi models, one row over from the Kimi For Coding membership already here. Its recharge-and-limits page (platform.kimi.ai/docs/pricing/limits, read 2026-09-05) opens with “you need to recharge at least $1 to start using, and when your cumulative recharge reaches $5, you will receive a” voucher, and adds that “Vouchers do not count towards the cumulative recharge total”; Tier0, the $1 tier, is 1 concurrent request, 3 RPM, 500,000 TPM and 1,500,000 TPD. No free balance, no free model, and the China platform (platform.moonshot.cn) is the same shape in yuan. platform.moonshot.ai now redirects to platform.kimi.ai, so all three apexes are here. Reopens if: A signup balance or a free tier with a figure on the pricing page.2026-09-05 — Still linked by two of the lists read on 2026-09-05. None of kluster.ai, platform.kluster.ai or api.kluster.ai resolves in DNS that day, and the last read of its successor lists recorded the API sunset as 2026-06-09. Reopens if: The domain resolves and a page carries a free tier.2026-09-05 — Honest and OpenAI-compatible by URL swap — its FAQ says “We offer free models, use them to test before making a purchase” — and the models page (read 2026-09-05) marks exactly one row FREE: MythoLite, a LLaMA-2 13B roleplay tune with a 2,560-token context and a 150-token output cap. Every coding-tagged row beside it (DeepSeek V4 Flash, GLM 4.7, GPT OSS 120B and 20B) is metered in credits. A free lane that cannot finish a function is not one a coding agent can use. Reopens if: A coding-capable model marked FREE on the models page.2026-09-05 — Reached through the models.dev digest (1/180 rows at cost 0, nvidia/nemotron-3.5-lightning-30b-a3b). merge.dev/pricing/gateway has a Free plan — “Test the core functionality before you upgrade”, “Access all free LLMs”, “Route requests automatically with built-in fallback”, “Start without a credit card” — and a Pro plan at “Pay LLM cost plus a 5% fee. Get $10 in free credits every month”, card required. Which LLMs are free is written nowhere readable: the docs’ Models page is 275 characters of navigation, the catalog at api-gateway.merge.dev answers 401 keyless, and the vendor-access page says organizations created after a vendor’s cutoff on Free or Pro get no access to that vendor “until access is granted”. A plan that names no model and no quota gives a probe nothing to hold, and the one zero-priced id this list knows of came from someone else’s key. Reopens if: Merge publishes which models the Free plan serves and at what limit, or opens its catalog keyless with prices.2026-09-05 — Reached through the models.dev digest (4/592 rows at cost 0). The free model is the web chat’s: nano-gpt.com/get-started says “Start chatting with a free model—no account or balance required”, and the API quickstart says “If you haven’t deposited yet, add some funds to your balance. Minimum deposit is just $1, or $0.10 when using crypto”. The keyless catalog at nano-gpt.com/api/v1/models?detailed=true prices 616 rows and only two at 0 per token — nano-gpt-help, the vendor’s support assistant, and auto-model, a router billed as whatever it routes to; the two Qwen3.5 Omni rows the digest counts as free carry “varies_by_modality” and a non-zero cost estimate. “API at list prices — no markup” is the offer, and it is a metered one. Reopens if: A chat model priced 0 in the detailed catalog that the docs say answers on a zero balance.2026-09-05 — A Chinese list carries it as a free LLM API. On 2026-09-05 the site renders as “nCompass — GPU performance engineering for opencode” and its pricing page carries no API rate card and no free tier — the inference product the list remembers is not what the site sells now. Reopens if: An inference API with a free tier appears on the site.2026-09-05 — Named by every list read on 2026-09-05, usually as “$50,000 in free credits”. The pricing page (developers.openai.com/api/docs/pricing) prices every model per token, marks one row Free — omni-moderation-latest — and carries no signup credit and no free tier. The one free thing is the help-center article “Sharing feedback, evaluation and fine-tuning data, and API inputs and outputs with OpenAI”, which answers 403 to this repository’s fetches and, as quoted by the developer forum and the third-party guides that cite it, says some organizations may qualify for daily complimentary tokens on traffic shared with OpenAI, that eligibility is shown only inside the organization’s data-sharing settings, that the tokens are usable only with a positive account balance, and that the shared prompts and outputs are used for training. An offer decided per account on a page no probe can read, unlocked by a prepaid balance and paid for with the prompts, is not a quota this list can publish. Reopens if: OpenAI states the eligibility rule and the daily amounts on a public page, or opens a free tier that needs no balance.2026-09-05 — Reached through the models.dev digest (3/137 rows at cost 0). The vendor’s own catalog at api.poe.com/v1/models is keyless and carries 347 rows, 146 of them with a pricing block and none at 0 — the three the digest counts as free, gemma-4-31b, kimi-k2.5-fw and gpt-5.3-codex-spark, are rows the catalog publishes no price for at all. creator.poe.com’s API page says who pays: “All Poe subscribers can use their existing subscription points with the API at no additional cost”, a 402 insufficient_credits comes back at “balance ≤ 0”, and the rate limit is 500 requests per minute. Whether the daily points a free Poe account gets in the app reach the API is stated nowhere the docs reach, and poe.com/pricing is a 342-character client-rendered shell. Reopens if: Poe’s docs say API calls draw on a free account’s daily points, with the figure, or its catalog prices a chat row at 0.2026-09-05 — ppio.com/pricing (read 2026-09-05) is a metered rate card in yuan — DeepSeek V3.2 ¥2 / ¥3 per million tokens, Qwen3.8 Flash ¥0.8 / ¥2.7, GLM 5.3 Flash ¥0.8 / ¥2.8, Kimi K2 ¥4 / ¥16 — with GPU spot discounts and a 50% batch-inference offer; no free quota, signup grant or free model is on it, and the sign-up is a mainland flow. Reopens if: A free quota with a figure on a vendor page.2026-09-05 — Code-edit models (apply, merge, search, rerank) that a Chinese list carries as a free API. relace.ai/pricing (read 2026-09-05) prices every one per million tokens — $1.00 / $3.00 on Repos, Jacq and relace-search, $0.80 / $1.20 on relace-apply-3, $0.05 in on relace-rank — and names no free tier, credit or plan. Reopens if: A free plan or signup credit with a figure on the pricing page.2026-09-05 — Kuaishou’s AI cloud, named by two Chinese lists. The home page (read 2026-09-05) sells KAT-Coder-Pro V2.5 and DeepSeek V4 Flash 0731 behind API/SDK access and offers only a student discount campaign — “高校师生专属｜模型定制低至5折”, to 2026-09-15, for mainland, Hong Kong, Macau and Taiwan faculty and students — with no free quota anywhere on it; the console is a login page. Reopens if: A free quota with a figure on a readable page.2026-09-05 — Named by two Chinese lists for its 协作奖励计划, the data-collaboration reward. The program page (docs.volcengine.com/docs/82379/1391869, updated 2026-08-31) is client-rendered and came back empty to this repository’s fetch on 2026-09-05; what could be read is the vendor’s 2025-11-20 notice as quoted by nodeseek and linux.do — the reward for individual users rose “由原本每日每模型的50万免费tokens奖励提升至200万”, granted as a next-day rebate of what was consumed, for accounts that pass real-name verification and authorize their prompts and outputs for model training — and a 2026-09-01 notice that 38 models left the program. A mainland real-name account and a training authorization are two walls this list does not carry, and the page that would say so is unreadable. Reopens if: A server-rendered program page, and a sign-up readable outside the mainland.2026-09-05 — A third-party dataset reports “$100/month of Serverless Inference credits on the Free plan”. The vendor’s usage-limits page (read 2026-09-05) says instead that “Serverless Inference credits come with Free, Pro, and Academic plans for a limited time” — no amount, no expiry — and lists “$100/month” as the Free tier’s default spending cap on pay-as-you-go, which a free account must turn on once the credits are gone. The per-model price list (DeepSeek V4 Flash $0.14 / $0.28, Qwen3 Coder 480B $1.00 / $1.50 per million) carries no free row, and the service is “only available from supported geographic locations” under CoreWeave’s terms. Reopens if: A stated credit amount, or a model priced 0.2026-09-05 — Reached through the models.dev digest (1/1 row at cost 0, zdev). docs.zeldoc.ai: “Get an API key — Request an invitation and generate your key” — an invitation-only European enterprise platform (“Private. Governed. Sovereign.”) with ZDev as its own code model and ZRouter over “40+ models routed”; zeldoc.ai/pricing answers 404 and no page names a price, a quota or a free plan. Reopens if: A public sign-up and a published free tier.2026-09-05 — The mainland console of the vendor this list carries as Z.ai, named by five of the lists read on 2026-09-05 (one of them: “永久免费, 仅限速” — permanently free, rate-limited only — on the GLM Flash models). The same Flash models are the listed row’s free lane at api.z.ai, and what the China console adds is a mainland account whose sign-up this list cannot read. Here so the console stops being proposed as a second vendor. Reopens if: A readable sign-up page for non-mainland accounts, or a free model the international site does not carry.2026-09-02 — Reached through the models.dev digest, which lists api.nova.amazon.com with 2/2 rows at cost 0. Nothing here can be read: nova.amazon.com/dev, /dev/documentation, /dev/pricing and even /llms.txt all serve the same 3.8-kilobyte JavaScript shell with 13 characters of visible text, and api.nova.amazon.com/v1/models answers 500 to a keyless GET. Whatever the developer portal grants, no page states it to an ordinary HTTP client — the ModelScope bar, on the other side of the world. Reopens if: A server-rendered page states a free quota for the Nova API, or /v1/models answers.2026-09-02 — Reached through the codertesla/ai-coding-deals list, which credits it with an unquantified Community tier. The vendor’s pricing page carries two plans and neither is it: “BUSINESS $100 /month flat — no per-seat charge” with “$100/mo of usage included” for “up to 50 seats”, and “ENTERPRISE Custom”. The only trial on the page is a line in the SLA FAQ — “Trials and beta use: Community support only; our SLA does not apply to trials or beta usage” — with no terms, quota or sign-up path beside it, and the docs’ token-based pricing page prices every model at public API list price plus a 40% service fee from the first token. The home page names nothing free either. Should a Community plan return, its terms are still published (v1.9, 2026-01-14): “you grant us the right to use your Customer Code, Output, and Usage Data to train the artificial intelligence models”, against the paid page’s “No AI training allowed”. Reopens if: A free or community plan with a stated quota appears on the pricing page.2026-09-02 — A free account is not a free tier. crof.ai/home sells the on-ramp — “Get an API key — free account, no credit card required” — and crof.ai/pricing states the offer in one line: “Pay for what you use. Per-token pricing across every model we serve. No subscriptions, no minimums — only the tokens you actually burn.” The catalog behind it, ai.nahcrof.com/v1/models, is public and prices all 23 models; not one is at zero (the cheapest row read on 2026-08-30 was $0.35 in / $0.80 out per 1M). Reached from OmniRoute’s provider table, where it is filed under paid. Re-read 2026-09-02: still 23 rows and none at zero, the cheapest still $0.35 in / $0.80 out per 1M, and the only “free” on the pricing page is a JavaScript comment about the billing portal. Reopens if: A model priced at 0 appears in ai.nahcrof.com/v1/models, or the pricing page starts naming a recurring free allowance.2026-09-02 — Reached through the models.dev digest (3/56 rows at cost 0). The keyless catalog at api.empiriolabs.ai/v1/models prices 78 of 177 rows at 0 per token, and almost all of them charge by another unit — video, music, image, speech, search — while two chat rows, gemma-3-27b and mistral-small-3-1, print “$0.0040 Per Message” beside their zero token price. Three GLM Flash rows (glm-4-7-flash, glm-4-5-flash, glm-4-6v-flash) are the only ones whose pricing rows read “Input Free / Output Free”, with the note “Base token use is free” — the same models Z.ai serves free from its own gateway, which this list already carries. Getting Started lists “prepaid credits on your account (top up via the dashboard Billing page)” among the prerequisites, Billing says “EmpirioLabs AI uses a prepaid balance system”, and nothing read says a zero-balance account can call the Free rows. Limits are 50 RPM and 2M TPM per account. Reopens if: The docs say the Free rows run on an account with no top-up, or a sign-up credit is named.2026-09-02 — Reached through the models.dev digest, which lists 23 duo-chat rows at cost 0 because the price is not on the row. The product page, docs.gitlab.com/user/duo_agent_platform/, files the platform under “Tier: Premium, Ultimate” and says of the one place a free account appears, “Features available on the Free tier require the purchase of GitLab Credits” — the generally available features “consume GitLab Credits when used”. A credit you must buy is not a free tier. Reopens if: The docs state a credit allowance included with the Free tier at no charge, with a figure.2026-09-02 — Reached through the models.dev digest (12/12 rows at cost 0). The doc link the catalog carries, model.inferx.net/endpoints, answers 404, and the api url beside it answers 401 with “service failure: no tenant context: set a default tenant” to a keyless /v1/models. Nothing here names an offer, a price or a quota. Reopens if: A readable page states what a new account may call for free, and the catalog answers.2026-09-02 — Reached through the models.dev digest, which lists api.kimi.com/coding with 4/4 rows at cost 0 — the price is the membership, not the row. The product docs say “Subscribers can also obtain an API Key” and give the billing as “Membership subscription, monthly/annual payment, with rate limiting” or “Pay-as-you-go, top up and use”. The doc url the catalog carries answers 404. No free membership is named anywhere read. Reopens if: A free membership tier with API access, with its quota stated.2026-09-02 — Reached through the models.dev digest (2/182 rows at cost 0). The word “Free” names the platform plan, not inference: the pricing page’s own JSON-LD offer reads “Free … Access all 200+ models with a flat 5% platform fee on credit purchases, or bring your own provider keys for free”, the page says “Start free with no credit card. Pay only for what you use”, and the docs put it as “Pay per-token with prepaid credits at provider list rates, or bring your own provider keys (BYOK) for free”. The keyless catalog at api.llmgateway.io/v1/models prices 42 of 252 rows at 0 and every one of them is a router (custom, auto) or a media model billed by another unit — TTS, image, video. No text model is priced 0 and no sign-up credit is named. Open source under AGPLv3, which is a reason to self-host, not a free lane. Reopens if: A text model priced 0 in the catalog, or a sign-up credit named on the pricing page.2026-09-02 — Reached through the models.dev digest (6/6 rows at cost 0). The free plan is your own hardware: pendra.ai/pricing offers “Free £0 /month — For personal use — sovereign inference on your hardware. 1 self-hosted Pendra Worker”, the home page says “Bring your own GPUs with a free account, or let us manage everything”, and the paid Pro tier at £99 a month is what carries “£99 of credit to spend on on-demand GPUs”. api.pendra.ai/api/v1/models answers 401 to a keyless GET. A gateway to a GPU you supply is not free inference. Reopens if: A hosted model callable on the free plan without supplying a worker, with a quota on the pricing page.2026-09-02 — The word “Free” here names a rate-limit tier, not free inference. platform.publicai.co/plans puts four tiers side by side and the Free one buys “100 requests/minute — default when you create an API key”; the same page keeps the money in a separate paragraph — “Rate limits and token usage billing are separate … Token usage is billed against wallet credits”, with “New accounts receive starter credits when they first use billing features” and no figure attached to those credits anywhere on the portal. The catalog agrees: every one of the eleven models on platform.publicai.co/models carries a price (swiss-ai/apertus-v1.5-8b at $0.10/$0.20 per 1M, up to $0.82/$2.92 for the 70b), and none is listed at zero. An unnamed starter credit is the SiliconFlow reasoning again: a sign-up courtesy is not a free tier. The nonprofit itself is exactly what it says — a public-AI gateway fronting Apertus, SEA-LION, Olmo, Bielik and EuroLLM, and this list already carries SEA-LION from its own vendor. Re-read 2026-09-02: the plans page still says “every new API key starts on the free tier” and “new accounts receive starter credits when they first use billing features” with no figure attached, the models page still prices every row ($0.10 to $2.92 per 1M), and api.publicai.co/v1/models answers 401 to a keyless GET, so the catalog itself cannot be read without an account. Reopens if: The portal names the starter credit, or platform.publicai.co/models lists a model at zero.2026-08-30 — OmniRoute credits it with 22 free models and ~7M tokens a month; the vendor publishes nothing of the kind. www.bluesminds.com is an enterprise page whose only two calls to action are “Sign in” and “Request access”, /pricing 404s, and there is no docs host to read. The one thing its API does say is what it runs on: api.bluesminds.com/v1/models answers {“error”:{“message”:”Invalid token …”,”type”:”new_api_error”}}, the error shape of the New API resale panel, which is the software a key-pooling gateway is usually built from. Nothing here is verifiable and the free claim is a third party’s. Reopens if: BluesMinds publishes a pricing or docs page an ordinary HTTP client can read, and that page names a free allowance of its own.2026-08-30 — The wrong shape rather than the wrong price. Liquid’s product is the LEAP edge platform — “Discover, specialize, and deploy models on any device” — whose models run on the phone or laptop you ship them to, not behind a hosted endpoint an agent can call: api.liquid.ai/v1/models does not exist (404 from the marketing app) and labs.liquid.ai serves no catalog either. A local runtime is out of scope here for the same reason Ollama’s desktop build is, and the “keyless” row OmniRoute files for it names no endpoint. Reopens if: Liquid opens a hosted LFM API with a free lane a developer can call from outside its own SDK.2026-08-30 — Moved off the blocklist because the fact it rested on is not what the catalog says. The reason read “163 models, 77 of them paid frontier ids — claude-opus-5, claude-opus-4.8, the gpt-5.6 line, o3 — offered on a free plan of 150K tokens/day”, and api.navy/v1/models carries a premium flag per row: 50 rows are premium and they are precisely that frontier line (gpt-5.6-sol/terra/luna, 5.5, 5.4, 5.3-codex), while the frontier-named rows that are NOT premium are the mini, nano and lite tiers. There is also a token_multiplier per row — 20 for claude-opus-5, 16 for gpt-5.6-sol, 1 for the ordinary rows — so the quota is spent at a rate that matches the model’s cost. That is the shape of a metered plan, not of capacity given away. Not listed because nothing here can be read: api.navy and its docs answer a Cloudflare challenge to every non-browser client, so the free plan’s terms cannot be verified from outside a browser. Reopens if: NavyAI serves a pricing or docs page to an ordinary HTTP client, and that page states what the free plan grants and which models it covers.2026-08-30 — Moved off the blocklist, where it sat since 2026-08-11 with a reason that ended “delete this line if an account shows otherwise” — a verdict written to expire, filed where nothing expires. The finding itself still holds and is a watchlist finding: the entry once claimed “$10 free credits at the start of every month” on the strength of one dated announcement post, and no live vendor page corroborates a recurring grant. platform.reka.ai/pricing today sells Pay as you go at $0/mo with “Credits — Top up anytime”, and the only free thing it names is in the past tense, a banner reading “Free credits used”. A legitimate vendor with no free tier a developer can reach is the definition of this file, not of the blocklist. Reopens if: A Reka page states a recurring free grant, or the signup credit is named with a figure.2026-08-30 — Blocked on 2026-08-30 and un-blocked the same day for the same error as xKiro: the frontier ids that triggered it are the paid catalog, not the free lane. Of 227 :free ids only nine are frontier-named, and all nine are Google’s cheap tier (gemini-flash-lite, gemini-3.5/3.6-flash-lite, embeddings, robotics previews) — no Claude, no GPT. What keeps it off the list is what cannot be read rather than what was found: its catalog publishes no prices at all, so no probe here can tell a free id from a metered one; its 338 rows are owned_by: "custom" on 318 of them; and 17 Claude ids are spelled in a vocabulary no public catalog uses — claude-opus-5[1m], claude-sonnet-5-thinking[1m] — which a router may well have invented for a context-and-reasoning preset, but which nothing on the site explains. Its paid plans sell $50 of credit for $25, permanently. Reopens if: Its catalog starts publishing per-model prices, or a page states what the free lane grants and where the [1m] ids come from.2026-08-14 — Carried elsewhere as “a recurring ~10M tokens/month free allocation (no card)”. Its own pricing page says otherwise, twice on one screen: the heading “Hobbyist (free)” is followed by “Just getting started? Start the Hobbyist plan — 3-day free trial, then $5/mo”, and the model table under it is titled “Hobbyist $5/mo and up … Token usage billed per rates below” with a real price on every row (Qwen3 Coder 30B $0.04/$0.04, GPT OSS 120B $0.42/$0.90, Gemma 4 31B $0.99/$1.49, GLM-4.7 Preview $1.25/$2.75). So the free part is a three-day trial on a paid plan that also meters tokens, and the words “free tier”, “10M” and “no credit card” appear nowhere on the page. Its keyless api.ainative.studio/api/v1/models does answer 200 with 84 rows, but they carry no prices and include other vendors’ catalogs (claude-3-sonnet owned_by anthropic, arcee-trinity owned_by digitalocean) tagged source “playground”. Reopens if: AINative publishes a recurring no-card allowance on a served page, with the models it covers.2026-08-14 — The free plan is real, published and server-rendered: anyapi.ai/pricing reads “Free / Getting started / $0 / mo”, “ANY Tokens: 100K / day”, “Core features: All free and basic models”, “No credit card required”, and docs.anyapi.ai repeats it — “$0/month 100 000 ₳nyTokens per day for free / Basic and free models access” and “Free: Your 100K daily ₳nytokens reset every 24 hours”. What stops it is that nothing says what the quota buys. The plan cards count ₳nyTokens; the FAQ on the same pricing page defines a different unit and never reconciles them — “A credit is our universal currency for API usage across all 400+ models. One credit equals roughly 1,000 tokens for most models, but premium models like GPT-4 or Claude 3 Opus consume more credits per token”. At 1,000 tokens a unit the free plan would be 100M model tokens a day, and the $19.90 plan 100 billion a month, which cannot be what is meant. Which models count as “free and basic” is published nowhere, api.anyapi.ai/v1/models answers 401, and there is no per-model price page — so there is nothing to probe that ties a model to the tier, and no arithmetic to tell a real gateway from a resold pool. The pricing copy still headlines GPT-4 and Claude 3 Opus in 2026, which is its own answer about how current it is. Reopens if: AnyAPI publishes which models the free plan serves, or a keyless catalog, or defines its token unit against real model tokens.2026-08-14 — Not an inference free tier — GPU capacity by the hour, advertised as “Available now Badgr 4090”. The pricing page’s $0 is a subscription that does not exist: “Pay as you go $0 /month + usage: every account gets the same features”, “no setup fee, no monthly minimum”. A wallet, and not even a wallet with anything in it. Reopens if: Badgr starts granting inference credits, or serves models rather than renting the cards they run on.2026-08-14 — Its pricing FAQ says only that “new Baseten accounts come with credits so you can get to know the UI and experiment with deployments for free” — no figure, anywhere on the page. There is nothing a probe could anchor on, and a sentence with no number in it reads the same the day the credit stops being granted. Reopens if: Baseten publishes the size of the signup credit.2026-08-14 — Two tiers on its own pricing page and neither is free: “PAY AS YOU GO”, metered per token against prepaid credits, and “ENTERPRISE”, “Your model, your keys, your contract”. The $0 in the first column is the monthly commitment — “$0 / MO COMMIT … Committed spend None — pay for what you use” — not a token allowance, which is the same shape this repository rejected for AIHubMix’s neighbours and for ZenMux. The 300+ model table below prices every row. Reopens if: Blackbox publishes a free tier with a token or request allowance, on the pricing page rather than in an extension listing.2026-08-14 — A real bundled free tier — bolt.new/pricing gives the $0 plan “300K tokens daily limit, 1M tokens per month” — spent inside a hosted app builder. It buys generations in Bolt’s own workspace, not model access a developer can point at a repository they already have, and there is no client to install and no endpoint to call. That is the line this list draws: an entry has to hand you model access you can aim at your own code. Reopens if: Bolt ships a CLI or an API whose free plan includes the model usage.2026-08-14 — The wrong shape twice over. It is a serverless platform you deploy your own container to rather than a model endpoint you can aim an agent at, which is the BYOK case CONTRIBUTING excludes; and the free thing on cerebrium.ai/pricing is the Hobby plan’s allowance of seats and slots — “Up to 3 deployed apps”, 3 seats, 1 day of log retention — while the compute underneath is billed throughout, the page’s own worked example putting 500,000 requests at “roughly $309” a month. No $30 credit appears anywhere on it. Reopens if: Cerebrium starts serving hosted models on a free allowance rather than hosting yours.2026-08-14 — Re-confirms the 2026-08-06 reading, which was never written down. Its pricing page states the model outright — “Pay per token. No subscription, no minimum, no markup” — and the cheapest thing on it is a $10-a-month Plus plan bundling a daily request quota. The free-trial FAQ entry is a collapsed client-rendered answer with nothing behind it in the HTML. Reopens if: Chutes publishes a zero-priced model or a no-card trial quota.2026-08-14 — Nothing here can be read, so nothing can be verified: www.coze.com answers every path tried with the same 43,754-byte SPA shell — the API overview, the pricing guide, /robots.txt and /sitemap.xml alike — and the document API behind that shell replies {“code”:700012006,”msg”:”Login verification is invalid”}. No quota, model list or terms are served anywhere a probe could anchor. The shape is doubtful as well — its own og:title calls it an “AI Agent Intelligent Office Platform” and its API is /v3/chat against bots built inside its workspace, which is the hosted-app-builder case CONTRIBUTING excludes — but the verification rule alone settles it, the way it settled Morph LLM. Reopens if: Coze serves its developer docs or pricing to an ordinary HTTP client, and that page describes model access rather than bot invocations.2026-08-14 — Renamed since gpt4free wrote its adapter — whiterabbitneo.com now redirects to deephat.ai, “Deep Hat, formerly Whiterabbitneo, is Kindo’s uncensored cybersecurity model”. The four “Use it for free” and “Try it for free” buttons on that page all lead to the hosted app; the API route on the same page is “Get a demo”. The weights are on Hugging Face, which is a licence rather than a free tier, and is not an endpoint anyone can point an agent at without hosting it themselves. Reopens if: Deep Hat publishes an API with a free lane, rather than a hosted app and open weights.2026-08-14 — deepinfra.com/pricing serves 17,500 characters of text and not one of them is “free” or “credit” — the whole page is per-token rates for language models and per-execution-time rates for everything else. The signup credit it is still credited with elsewhere is not on the vendor’s own page. Reopens if: DeepInfra publishes a signup credit or a free model on its pricing page.2026-08-14 — Re-checked against api-docs.deepseek.com/quick_start/pricing, which is server-rendered and prices every model — $0.14 in / $0.28 out per 1M for the cache-miss tier, no zero row anywhere. This confirms the 2026-07-27 removal: the 5M-token signup grant that used to be cited for DeepSeek was only ever reported by third parties and appears on no official page. Reopens if: An official DeepSeek page states a free grant or a zero-priced model.2026-08-14 — Blocklisted earlier the same day on a reason that turned out to be wrong, and moved here rather than left standing. The claim was that its model ids anonymise the upstream they route to; dxnt.com/byok refutes it in its own words — “在控制台添加自有端点，用 {身份码}/{模型名} 调用。请求直接转发到你的供应商” — so the opaque four-character prefixes (ot5k/, its0/, ry5b/) are per-user identity codes for a customer’s OWN endpoint, not a hidden supplier, and dxnt’s own catalog names its models plainly as dxnt/qwen3.5-397b-a17b and the like. Its free route is four models at multiplier 0, mostly open-weight (nemotron-3-nano, llama-3.2-90b-vision, step-3.7-flash, sensenova-6.7-flash-lite), which is capacity anyone may host. What is left is not enough to block and not enough to list: the service calls itself a 中转网关, a relay gateway, discloses no sourcing for the proprietary Chinese models it relays, and its footer links 服务条款 and 隐私政策 both resolve to a page containing neither, so there are no terms and no legal entity behind any of it. Reopens if: DX TOKEN publishes terms of service and a legal entity, or states where the models on its own dxnt/ prefix are licensed from.2026-08-14 — A coding harness that brings no model usage of its own. Its installation docs say ForgeCode “needs access to at least one AI model”, walk you through entering an API key, and recommend reusing a ChatGPT Plus or Claude subscription you already pay for; forgecode.dev/pricing is a 404. The “10K tokens a day via OpenRouter” credited to it elsewhere is OpenRouter’s own free tier reached with the reader’s key, and OpenRouter is already an entry here. Reopens if: ForgeCode Services bundles model usage on a plan that costs nothing.2026-08-14 — friendli.ai/pricing prices everything it serves: per-token rates for the Model APIs — GLM-5.2 at $1.4 in / $4.4 out per 1M, gemma-4-31B-it at $0.14 / $0.4 — and per-second GPU time for dedicated endpoints, $2.9 an hour on an A100 up to $12.0 on a B300. The word “free” does not occur in the served text of that page at all, and its own “How do I get started?” answer is sign up, generate a key, make a request, with no grant in between. The $10 signup credit third-party lists quote is on none of it. Reopens if: FriendliAI states a signup grant or a zero-priced model on a page it serves.2026-08-14 — The largest free grant this repository has ever verified, and it still cannot be listed. Sber’s own tariff page, updated 22 July 2026, says “В рамках Freemium-режима пользователи получают 365 000 000 бесплатных токенов для генерации текста” and breaks it down in a served table that glues each figure to the ids it applies to — 250 000 000 for Lite (GigaChat, GigaChat-2), 40 000 000 for Pro, 25 000 000 for Max, 50 000 000 for GigaChat-3-Ultra — each valid 12 months, “Генерация текста выполняется в одном потоке”, one stream at a time. Every new account gets the Freemium tariff by default and no card is asked for; only the paid packages take one. The endpoint is OpenAI-shaped and printed in the docs: base_url “https://api.giga.chat/v1”, model “GigaChat”. What blocks it is the transport. Both api.giga.chat and gigachat.devices.sberbank.ru serve certificates issued by “C = RU, O = The Ministry of Digital Development and Communications, CN = Russian Trusted Sub CA”, chaining to a Russian Trusted Root CA that is in no mainstream trust store — measured here, not inferred: httpx fails both with CERTIFICATE_VERIFY_FAILED, and Sber documents the fix as installing the НУЦ Минцифры certificates at application or OS level. That is a change to the reader’s machine, and to what their machine trusts for every other TLS connection, which is not a thing a row in a table can ask for in a footnote. Registration is a second gate on top: the console takes a Сбер ID and issues scope GIGACHAT_API_PERS. Reopens if: api.giga.chat serves a certificate from a publicly trusted CA, or Sber documents an endpoint that does. At that point only the Сбер ID gate is left, which is the SenseNova case and lives inside an entry, and the anchor is already measured: the 365 000 000 sentence and the per-tier table sit in served HTML with ordinary spaces.2026-08-14 — The service does not answer. Both the site and its OpenAI-compatible endpoint return Cloudflare 522 “Connection timed out” — the origin is down, not the edge. An offer that cannot be reached cannot be verified, and a dead host is not evidence of a withdrawn tier either, which is why this is a watch and not a block. Reopens if: The origin answers again and its free lane is still there.2026-08-14 — The business is GPU rental now, not a free inference lane — hyperbolic.ai/pricing is a 404 and the home page sells on-demand GPUs “billed by the hour” and reserved capacity at discounted rates. The $1 inference credit third-party lists still quote is not on any page it serves. Both domains are recorded because the lists still point at the older one: hyperbolic.xyz/pricing and app.hyperbolic.xyz redirect to the .ai host (measured 2026-08-14), so a verdict filed under one name would have left the other free to be re-proposed. Reopens if: Hyperbolic publishes an inference free tier or signup credit again.2026-08-14 — The free model is real and the vendor says so on its own page: the HTTP docs describe Lite as “轻量级大语言模型 … 具有更高的响应速度，支持免费使用” (a lightweight model, faster to respond, free to use), give the model id as lite and publish an OpenAI-compatible base url, https://spark-api-open.xf-yun.com/v1/. An anchor exists too, contrary to an earlier note here: 「具有更高的响应速度,支持免费使用」 sits in the Lite column of that page’s version table in served HTML, glued to the id by 「lite指向Lite版本」 beside it. What blocks the listing is reachability. No served page states a quota, a QPS or a duration for the Lite lane; the figures that exist — the 0元/百万tokens cell and the 200,000-token free package — appear only after JavaScript runs on xinghuo.xfyun.cn/sparkapi, and that same table conditions the package on 个人认证 or 企业认证, real-name verification. Whether lite itself is reachable without it is stated nowhere, and console.xfyun.cn renders 「加载中……」 and nothing else. That is the ModelScope case: a real free model this list cannot yet tell a developer how to reach. Reopens if: iFlytek states, on a served page, what the Lite lane costs and whether it needs 实名认证 — or the registration gate turns out not to need a mainland-China identity.2026-08-14 — The $0 plan is a gateway and observability tier, not free inference: it lists “1M included Gateway requests, 1M tracing spans, 14-day data retention, 1 seat, 30 req/min” and is priced “$0 + usage”, so the tokens themselves are still bought. The only opening credit on the page, $50, is attached to the $250-a-month Growth plan. Reopens if: The $0 plan includes inference credits or a zero-priced model rather than routing quota.2026-08-14 — The docs are readable after all, at internlm.intern-ai.org.cn/doc/docs/ — and the trailing slash is load-bearing, since the same path without it answers a 168-byte stub, which is what an earlier check here mistook for an empty domain. 197KB of served HTML confirms the base url chat.intern-ai.org.cn/api/v1, the four model ids, an Anthropic-format endpoint with a Claude Code guide, and a quota: 「API 流控限制：默认每用户 每分钟限制30次」, thirty requests a minute rather than the ten a feed credits it with. What is missing is the offer itself. Across all nine doc pages the words 免费, 收费, 付费, 计费 and 价格 do not occur once: the service states no price, free or otherwise, so there is nothing to verify and nothing an anchor could hang on — 每分钟限制30次 is a rate limit that would read exactly the same the day it started billing. Reopens if: Intern AI states a price, or states that the API is free, on a served page.2026-08-14 — Listed here from 2026-07-22 for a recurring 100K free tokens a day, and delisted because that quota is published nowhere on Meituan’s own site today. Two dated entries in its change log explain what happened. “Version: 2026-05-29 — LongCat Model Service Sunset” retired the whole Flash line the free quota applied to: “Effective May 29, 2026, the platform will retire the following 6 models: LongCat-Flash-Chat, LongCat-Flash-Thinking, LongCat-Flash-Thinking-2601, LongCat-Flash-Lite, LongCat-Flash-Omni-2603, LongCat-Flash-Chat-2602-Exp”. Then “Version: 2026-06-30 — LongCat-2.0 Release & Billing Now Available” turned on Token Packs and pay-as-you-go. The API is alive and good, but it is now priced: “Pay-As-You-Go currently supports LongCat-2.0” at $0.75 uncached input and $2.95 output per 1M tokens, and its prerequisites end with “Top up your balance at Billing → Recharge”. The words “free” and 免费 appear nowhere in Quick Start, Token Pack, API Pay-As-You-Go, the FAQ, the API overview or the pricing page — the one “free” in the change log is about chatting on the consumer LongCat Chat app, not the API. The platform page itself is a 21-character client-rendered shell. Complimentary credits exist as a concept (“If you also have complimentary credits, those are consumed before your paid balance”) but nothing states that a new account is granted any, how many, or how often. Our probe never noticed, because it anchored on the endpoint path and the model name rather than on the offer — novita’s failure, again. Where our figure came from, for whoever is tempted to re-add it: free daily quotas at LongCat belong to the era before GA. The Chinese change log grants one on 2026-04-20, to the PREVIEW model — “初始额度5,000,000 Tokens/天 … 每日最多可获得120,000,000 Tokens” — and 赠送 appears nowhere in that log at all. The reference response for GET /openai/v1/models in the docs now lists exactly one entry, “id”: “LongCat-2.0”, “owned_by”: “LongCat”. Reopens if: LongCat publishes a free daily token quota, or a stated signup grant, on a server-rendered page of its own.2026-08-14 — Same hosted-builder shape as Bolt, and thinner: its own FAQ says “Paid plans have access to a credit balance” and describes the free plan only as having “grants included in the free plan”, with no figure in the served HTML. Reopens if: Lovable publishes a free-plan allowance in server-rendered HTML and a way to spend it outside its own workspace.2026-08-14 — Recurring and real — modal.com/pricing gives the $0/month Starter plan “$30 / month free credits” — but the credits buy compute, not inference. They are spent on containers and GPU-seconds running a model the developer deploys themselves, so there is no vendor in the transaction, nothing to probe behind an OpenAI-compatible URL and nothing to put in a Models column. It fails this list’s bar on shape, the way a free Colab GPU does. Reopens if: Modal serves models of its own behind an OpenAI-compatible endpoint on a free allowance.2026-08-14 — Rebranded from Nebius AI Studio — studio.nebius.com now redirects to tokenfactory.nebius.com. Its served pages carry no free tier at all: the pricing page prices storage and GPU-hours, the product page offers a “Start free” button and a sales contact, and the docs quickstart carries no grant, quota or trial. The catalog answers 401 without a key. Nothing here is a free offer a developer can name, so there is nothing to list. Reopens if: Token Factory publishes a signup grant or a free model lane on a served page.2026-08-14 — An OpenAI-compatible gateway with a keyless catalog — api.ofox.ai/v1/models answers 130 rows unauthenticated — and every text model in it is priced. Its eight GLM rows start at glm-4.7-flashx for $0.072 per M input, and there is no glm-4.7-flash in the catalog at all, so the third-party claim that it serves GLM-4.7-Flash at $0 per M is about a model it does not carry. The fourteen rows that do read prompt 0 and completion 0 are video, image and transcription models whose real price sits in a per-second or per-image field beside them. Reopens if: A text model in api.ofox.ai/v1/models prices both prompt and completion at zero.2026-08-14 — Pay-per-use with no free allowance of its own. The word “free” does not occur in the served text of docs.perplexity.ai/getting-started/pricing, which prices Sonar at $1 in and $1 out per 1M tokens, search requests at $5 per 1000 and each tool per invocation; the page’s only credit language is “Purchase API credits through AWS Marketplace”. Credits bundled with a paid subscription, if any, would be a wallet behind a paywall rather than a free tier. Reopens if: Perplexity publishes an API allowance reachable without a paid subscription or a credit purchase.2026-08-14 — Nothing could be verified, and that is the finding. www.phind.com and www.phind.com/plans both answer HTTP 403 to every client tried here, including one sending a full browser header set — so no probe on this domain can read an offer, and this list would be repeating a claim rather than checking one. Worth the record because the name is otherwise a plausible candidate: a developer-facing answer engine with an editor extension is exactly the shape CONTRIBUTING admits when the free usage is bundled. Note also that gpt4free does not call phind.com at all — its adapter points at phindai.org, an unrelated clone, now on the blocklist. Reopens if: Phind serves its plans page to a non-browser client, or publishes free terms anywhere a probe can reach.2026-08-14 — replicate.com/pricing says “You only pay for what you use on Replicate” and prices every model by hardware-second or by token. The only “free” in the served text is the “Try for free” sign-in button in the header — no signup grant, no recurring allowance, no zero-priced model. Reopens if: Replicate publishes a signup grant or a recurring allowance rather than pay-as-you-go only.2026-08-14 — The Free plan does carry an “API Access” checkmark, and the docs say what it is worth: “You can create a key before funding the account, but model requests will not succeed until the account can consume DIEM, bundled credits, or USD”. DIEM comes from staking VVV and bundled credits from a subscription, which venice.ai/pricing states from the other side — “Pro subscribers get free-tier API access”. So the free lane is a key that cannot call anything, and every row of the docs price table is quoted in dollars per million tokens. Reopens if: An unfunded Venice account gets a callable allowance, or a model appears priced at zero for the Free plan.2026-08-14 — The Free plan is real and costs nothing, but it bundles no model usage: warp.dev/pricing lists it as “Reload credits at pay-as-you-go rates” and “Bring your own AI inference”, while the 1,500 included credits begin on the $20 Build plan. Third-party lists still credit the free plan with 150 then 75 AI credits a month; that allowance is no longer on the vendor’s own page. A terminal that asks for your own key is the aider case, not the opencode one. Reopens if: The Free plan on warp.dev/pricing lists an included monthly credit or token allowance instead of pay-as-you-go reloads.2026-08-14 — docs.x.ai prices every surface — text, Agent, TTS, STT, Imagine — and names no free tier or signup grant. x.ai/api answers HTTP 403 to every non-browser client, so even a stated offer could not be probed from here. The free API credits xAI ran in exchange for data sharing are not on any current page. Reopens if: A free tier or signup grant appears on a page that answers a plain HTTP client.2026-08-14 — A separate surface from the archived MiMo Code entry, and also paid. The keyless platform.xiaomimimo.com/api/v1/models publishes five models and prices all of them, the cheapest being mimo-v2.5 at $0.14 in / $0.28 out per M. The Token Plan is a paid subscription — the site advertises MiMo Claw at 14.9 yuan a month as “stackable with Token Plan” — and both mimo.mi.com/pricing and platform.xiaomimimo.com/pricing are client- rendered shells serving 29 characters of text, so no probe could read an offer there even if one appeared. Reopens if: The keyless catalog carries a zero-priced row, or a free tier appears in server-rendered HTML.2026-08-14 — A genuinely free, genuinely keyless lane that is the wrong kind of thing. you.com/docs offers “Try it free — no API key required”: point any MCP-enabled tool at https://api.you.com/mcp?profile=free and “use you-search with no signup or credentials”, capped at 100 queries a day, with new accounts getting $100 in credits for the rest. It names Claude Code, Cursor, VS Code and JetBrains by name, so it is aimed squarely at the readers of this list — but what it hands them is web search, not model access, and CONTRIBUTING asks an entry for a lane you can point a coding agent’s model at. Reopens if: You.com opens a free lane on an inference endpoint rather than on search and retrieval — or this list decides that free agent tooling belongs in a section of its own, in which case this is the first row in it.2026-08-14 — Unreadable rather than absent: yupp.ai answers HTTP 429 to every request made here, browser headers included, and the docs.yupp.ai its adapter implies does not resolve at all. gpt4free’s own class for it carries working = False and needs cloudscraper, which is the same verdict from the other side. Nothing states an offer, so there is nothing to verify. Reopens if: Yupp serves a plan or docs page to an ordinary HTTP client, and that page describes an API rather than a chat product.2026-08-14 — The scout proposed this on 2026-08-14 and the probe passed, because zenmux.ai/api/v1/models prices six of its 156 rows at 0 — z-ai/glm-4.6v-flash-free, z-ai/glm-4.7-flash-free, deepseek/deepseek-v4-flash-free, inclusionai/ling-3.0-tiny and the two sapiens-ai/agnes-*-flash. A zero in a price column is not a lane you can reach: ZenMux’s own subscription guide says of its $0 Free tier “The Free plan can only be used in Studio Chat on the web (roughly 5 conversations per 5 hours) and provides no API Request access”, and the other billing system, Pay As You Go, is documented as “prepaid balance + pay-as-you-consume” with a $5 minimum top-up. Neither published path hands a developer an API key that calls those six ids without paying, so the free price is a discount inside a paid plan rather than a free tier. Reopens if: ZenMux documents that a zero-balance Pay As You Go key may call the -free ids, or the Free Builder tier gains API access.2026-08-05 — Legitimate commercial routing layer, and the third unrelated service to carry the TokenRouter name — check the TLD before acting on any lead here, because .com is the listed PaleBlueDot gateway and .me is blocklisted. Its quickstart says “You need at least one AI provider key to route requests” and sends you to platform.openai.com, i.e. BYOK-only with no bundled model usage, and its “Free tier available” covers the routing plan, not tokens — the pricing page’s own words are “Pay only for routing and insights”. Not blocklisted precisely because nothing is wrong with it. Reopens if: TokenRouter.io starts bundling model usage rather than routing only.⏰ — the verdict is older than 90 days, no longer suppresses anything, and is due for a fresh look.