Current check — 21 October 2026: Claude’s general startup cohort is now closed, with current-member access and a separate conditional partner-VC route. Cloudflare’s paid-model list now has eight entries, including
deepseek-v4.1-flash-1017. The dated observations below remain a record of the earlier quarter; the change log and linked offer pages carry the current terms.
Free LLM API access did not vanish in 2026, but the offers in our tracked sample became narrower. We re-verify every offer in our directory against the provider’s own pages at least every 30 days, and keep a public change log. Laid side by side, the changes from July to October point one way.
The ten changes
Every row is quoted from the provider’s own page; each links to the offer page or change-log entry carrying the source and the date we checked it.
| When | Provider | What changed | Provider’s words |
|---|---|---|---|
| 30 Jul 2026 | GitHub Models | Free inference API retired outright | ”As of July 30, 2026, GitHub Models has been fully retired.” |
| Aug 2026 | Google Cloud $300 | Credit stopped covering partner models such as Claude | ”You can’t access or use the $300 credit for a generative AI partner model…” |
| Aug–Sep 2026 | Cloudflare | 0 → 5 → 7 models need a paid billing method despite free Neurons | ”Some models require a paid billing method” |
| Sep 2026 | Cerebras | Free tier became a $5, 30-day trial that needs a card | ”Is there a permanently free tier? No.” |
| by Sep 2026 | Groq | Llama chat models gone from the free table; Compound shut down 21 Sep | ”Compound was decommissioned on September 21, 2026.” |
| by Oct 2026 | Hyperbolic | Serverless inference API and playground retired | ”The serverless inference API and the playground have been retired.” |
| 7 Oct 2026 | Hugging Face | Free accounts’ monthly credit: $0.10 → none | ”Free Users | None | - | yes (credits purchase required)“ |
| Oct 2026 | Mistral | ”$10 /mo in API credits” no longer published for the Free plan | Free mode now “lets you … use included monthly usage within the limits shown on the Limits page” |
| Oct 2026 | Fireworks AI | ”$1 in free credits” gone from its pricing page | ”Simply add a payment method to buy credits” |
| Oct 2026 | Together AI | Minimum purchase to use the platform: $5 → $15; still no free trial | ”Access to the Together platform requires a minimum $15 credit purchase.” |
Four of the ten remove a free allowance outright (GitHub Models, Hyperbolic’s serverless API, Hugging Face, Cerebras’ free tier). Three fence off part of one (Google’s partner-model exclusion, Cloudflare’s paid-only list, Groq’s model turnover). Two make a figure disappear (Mistral, Fireworks). One raises the price of entry (Together).
A second finding: “no card” is usually our assumption, not theirs
When we went back through our own directory this month, we found that most “no credit card” claims — ours included — were inferences. Of the eight ongoing free tiers we list, three providers say in their own words that no card is needed: OpenRouter (“The Free plan works without a payment method”), Mistral (“API access is enabled by default with no credit card required”) and Cloudflare (“Start building for free — no credit card required”). Google, Groq, Cohere, NVIDIA and Leonardo.Ai offer free tiers but do not say either way. We now label those “not stated” rather than “no card” — the inference has already failed once, when Cerebras added a payment step.
The counter-evidence
The trend is not uniform, and a fair reading has to include what went the other way:
- DeepInfra’s DeepStart grew from up to 1 billion to about 50 billion free tokens for funded startups.
- Anthropic’s Claude Startups opened to bootstrapped founders on 6 October 2026 — though within days Anthropic said it was over capacity on the $1,000 credit it had announced. That expansion is historical; the general cohort is now closed.
- Mistral’s Free mode now states verbatim that no card is needed, even as its figure disappeared.
- AWS says a payment method “isn’t required to sign up for most new customers” for its up-to-$200 Free Tier, a loosening.
- Groq’s free model that replaced Llama as our worked example in September allows 200,000 tokens a day, against the old model’s 100,000.
Note where the growth is: in startup programmes, which require an application, not in the self-serve free tiers anyone can use today.
How we know — and what we cannot claim
Method. Every offer in our directory is re-read against the provider’s own pages at least every 30 days. A figure is published only if it is a string the page renders to a user, and every changed figure is re-checked by a second, independent pass whose job is to refute it. Changes are recorded with dates in the change log; the full dataset, with each record’s source URL and last-verified date, is at /api/offers.json and documented on /data.
A one-day survey. On 19 September 2026 we checked 24 providers for a free tier or signup credit. Ten yielded an offer with a figure we could quote. Four contradicted themselves across their own pages, three described an offer without a figure, two had no such offer, and one (GitHub Models) had been retired. The remaining four were held back for reasons unrelated to their figures — a scope question, or a check we had not finished.
Limits — read these before quoting us.
- The sample is small and not random. It is the providers we track, which skews towards those that had free offers to begin with. It is not a market statistic.
- It is one quarter. Three months of direction is a pattern, not a law. Prices for running models are falling at the same time, and free allowances may re-expand.
- Dates are when we saw the change, except where a provider dates it (GitHub 30 July; Groq 21 September; Hugging Face’s docs change 7 October).
- We measure what providers publish, not what they grant. A provider can be generous in practice and vague on paper; Mistral may still include usage it no longer quantifies.
- What would falsify this: new free tiers launching, or several of these changes reversing, over the next two sweeps. We will report either.
What this means if you build on free tiers
Hold keys at several providers rather than depending on one, check the last-verified date before you build on a number, and treat a missing figure as a warning rather than a promise. The tiers that remain are genuinely useful — see the current list on free AI API tiers, and the full record of what moved, month by month, in the October briefing.