Free AI credits briefing — September 2026: what changed

Superseded — read the October 2026 briefing for current figures. This page is kept as a dated record of what was true on 13 September 2026. Several things below have since changed: Groq removed qwen/qwen3.6-27b from its free plan by 19 September and decommissioned its groq/compound systems on 21 September; Hugging Face ended the monthly credit it gave free accounts on 7 October; and of the six tiers this page calls no-card, three — Gemini, Groq and Cohere — never actually say so, which we corrected in October. AWS’s new-account credits are also not documented as covering Claude, contrary to what we wrote here. The change log records each change.

We re-verified all 15 offers in the directory against providers’ official pages on 13 September. Four material changes — and the biggest one is a correction to something we told you ourselves.

The big one: Cerebras is not a free tier any more

We listed Cerebras Inference as a no-card free tier worth up to 1M tokens/day per model. That listing is no longer accurate, and Cerebras’ own documentation now says so in plain terms. Its rate-limits page has gained an FAQ entry:

Is there a permanently free tier? No. The Free Trial is time- and credit-bounded: $5 in credits that expire 30 days after they’re granted. Once your credits are used up or expire, you need to purchase credits (Pay as You Go) to keep using public endpoints.

And the credits are no longer card-free:

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.

Three things follow, and we have applied all three across the site:

1. It requires a card. We previously published cardRequired: false for Cerebras, inferred structurally from the absence of a payment step in the signup flow. That inference was wrong, and Cerebras now contradicts it verbatim.
2. It is a trial credit, not a provider free tier. We have moved it to the trial-credit category, which also removes it from free AI API tiers and from the no-card lists in our guides.
3. The headline number changes. The 1M tokens/day figure still exists, but it is a rate limit inside the 30-day trial, not a recurring allowance. The honest headline is $5, expiring in 30 days.

This is the failure mode the site exists to catch, and it caught our own listing. The August briefing recorded the $5 credit as a bonus on top of an ongoing free tier, and that reading has not survived contact with Cerebras’ current docs. The per-model limits themselves are unchanged (5 requests/min, 30K uncached tokens/min, 90K total tokens/min, 1M tokens/hour, 1M tokens/day); the documented models rotated again, from gpt-oss-120b / zai-glm-4.7 / gemma-4-31b in August to gpt-oss-120b and qwen-3.8-27b now.

Groq’s free plan has no Llama chat models left

Groq’s free plan is still a genuine no-card free tier with the most forgiving per-minute limit we track — but the model list underneath it has turned over completely. The published Free Plan table no longer contains llama-3.3-70b-versatile or llama-3.1-8b-instant, the two models every one of our worked examples used.

The free-plan chat models are now:

Model	Requests/min	Requests/day	Tokens/min	Tokens/day
openai/gpt-oss-120b	30	1,000	8,000	200,000
openai/gpt-oss-20b	30	1,000	8,000	200,000
qwen/qwen3.6-27b	30	1,000	8,000	200,000
qwen/qwen3.8-27b	30	1,000	8,000	200,000
groq/compound	30	250	70,000	—
groq/compound-mini	30	250	70,000	—

Two practical consequences. The daily cap is now uniform at 1,000 requests for every free chat model, so the old advice to pick the 8B model for its 14,400 requests/day no longer applies — that headroom is gone. And the tokens/day ceiling doubled, from 100,000 on the old 70B example to 200,000 on gpt-oss-120b, so token-heavy work actually got more room while request-heavy work got less. The only entries still showing 14,400 requests/day are the llama-prompt-guard-2 safety classifiers, which are moderation models rather than chat models, so we do not present them as an option.

We have replaced the Groq figures in every guide that quoted them.

Cloudflare: seven models now need a paid billing method, not five

Cloudflare Workers AI still gives “a total of 10,000 Neurons per day at no charge”, resetting at 00:00 UTC, at $0.011 per 1,000 Neurons beyond that — all unchanged. The exclusion list grew: @cf/zai-org/glm-5.3 and @cf/zai-org/glm-5.3-flash joined kimi-k2.6, kimi-k2.7-code, glm-5.2, deepseek-v4-flash-0731 and deepseek-v4-pro-0813 on the list of models that “require a paid billing method” regardless of your free Neurons. The pattern is now clear enough to plan around: the large third-party models arrive behind billing, and Cloudflare’s own and smaller open models stay in the free allowance.

GitHub Copilot Student sign-ups have reopened

When we published the GitHub Student Developer Pack in August, education.github.com/pack still displayed “New plan sign-ups are temporarily paused as we ensure a high-quality experience”, contradicting GitHub’s own changelog, which had announced a gradual reopening. We published the Azure half only and described the Copilot half without quantifying it.

That banner is now gone. The pack page states that “GitHub Copilot Student is available to verified students”, with “unlimited code completions and an allowance of GitHub AI Credits, plus limited chat and agent usage with models available through auto model selection only”. So Copilot Student is claimable again — but note what is still missing: GitHub publishes no number for that AI-credit allowance, so we describe it and do not quantify it. The offer’s credit figure stays scoped to the verified $100 Azure credit, which the pack page also now explicitly labels “no credit card required”.

Re-confirmed unchanged

Eleven offers re-verified with no change to any published figure:

• Gemini API free tier — free tier across the Gemini line-up, still “free of charge in all available regions”, still no published numeric limits.
• Groq — the tier itself: still free, still no card, still 30 requests/min. Only the model list changed (above).
• OpenRouter — 20/min, 50 :free requests/day, 1,000/day after a $10 lifetime purchase.
• Cohere — 1,000 API calls a month, Chat at 20/min.
• Hugging Face — “$0.10, subject to change” monthly, $2.00 on PRO. Checked on the ≤2-week cadence it is held to.
• Google Cloud $300 — 90 days, and both generative-AI exclusions still stated verbatim, so it still cannot fund Claude on Vertex AI. AWS Bedrock remains the documented cloud route to Claude.
• AWS Free Tier — up to $200, 6-month plan, 12-month credit expiry, card required.
• AWS Activate — Founders up to $5,000, Portfolio up to $200,000, pre-Series B and founded within 10 years.
• Azure free account — $200 for 30 days plus 12 months of free services.
• Microsoft for Startups — $200 at setup, up to $150,000 on approval, pre-seed through Series C. Microsoft re-dated this page on 11 September, so it is freshly maintained rather than merely unchanged.
• Google for Startups and Claude for Startups — credit structures and eligibility wording identical to August.

Two long-running questions, one answered

AWS Bedrock’s Free Tier listing is confirmed again. In August we could not read the AWS Free Tier offers catalogue — it appeared to our tooling as an empty client-side widget — so we stopped asserting that Bedrock was covered, without claiming it had been removed. That caution was right, and the answer is now readable: the catalogue lists Amazon Bedrock as “Available on both plans”, with the expanded note “Use credits to access Amazon Bedrock features in both Free and Paid plans”. Nothing had changed; only our ability to see it had. Bedrock AgentCore remains “Paid plan exclusive”.

Google’s two startup properties still disagree about company age. startup.google.com states “Founded within 5 years of applying to the program (or 10 years for Scale)”. cloud.google.com/startup states no age criterion at all and frames eligibility purely by funding stage. Since the two official sources disagree, we continue to publish only the funding-stage criteria they agree on.

Still not listed, and why

• OpenAI — no general free API tier. Unchanged since we started tracking.
• Anthropic’s first-party API — the conflict is now four sweeps old and completely static. The pricing FAQ says “New users receive a small amount of free credits to test the API” with no amount, no card requirement and no expiry, while the support documentation says “Buy credits before you use the API”. Only an authenticated Console session could settle it, so we have moved this one to an annual re-check rather than re-reading two unchanged pages every month. We publish no Anthropic signup credit.
• Mistral — La Plateforme’s free mode exists, but its limits live behind a login and no numeric limit, card requirement or verification requirement is quotable from a public page. Still held as a draft.

The takeaway

Two of this month’s four changes moved in the same direction: a free tier became a trial (Cerebras) and a free model list shrank (Groq’s Llama models, Cloudflare’s two extra exclusions). That is worth watching. The no-card free tier is not disappearing — six of them are intact and Groq’s actually got more generous per token — but the pattern of the last two sweeps is that the most expensive models drift behind a billing method first, and the tier framing follows.

The practical advice is unchanged and now better supported: hold keys at several providers rather than depending on one, and check the last-verified date before you build anything on a number.