Free AI compute in 2026: every free tier, trial, startup grant, and student deal — verified

There is far more free AI compute available in 2026 than any single “free LLM API” list captures — but it’s spread across five very different rungs, and most roundups mash them together, mislabel the units, and quote numbers that changed months ago. This is the full map: free-forever API tiers, one-time trial credits, student deals, startup model credits, and six-figure cloud grants. Every figure below was verified against the provider’s own page on 2026-06-26, with the proof links to match — and where a number is a third-party guess and not an official one, this post says so instead of pretending.

Free AI compute is a ladder, not a list

The single biggest mistake in every “get free AI” article is treating one bucket as the whole picture. A hobbyist who wants to call an API for a side project and a funded founder who wants $200k of GPU credits are not shopping in the same place — and the offer that’s perfect for one is useless or unreachable for the other.

So before the directory, the structure. Free AI compute sorts into five rungs, each with a different gate and a different ceiling:

The rungs are roughly ordered by how hard the gate is. Rung 1 asks for an email. Rung 5 asks for a cap table. Work out which rung you actually qualify for first — it saves you from applying to a $150k program you’ll be rejected from, or overlooking the email-only tier that would have covered your whole project.

Rung 1 — Free-forever API tiers (no card, no expiry)

This is the rung most people mean by “free AI.” You sign up, you get an API key, and you can call models at $0 indefinitely within rate limits. No balance to run out, no expiry. Ten providers genuinely qualify in 2026 — here they are with the limits as verified today:

A few of these need their footnotes spelled out, because the headline number alone will mislead you:

Google AI Studio is the strongest free tier for capability — you get Gemini 3.5 Flash and Gemini 2.5 Flash-Lite with the full 1M-token context window and no card. Google’s own rate-limits page now defers the exact per-minute and per-day numbers to your live AI Studio dashboard (they vary by model and tier), so treat the ~1,500 requests/day figure as the established ballpark, not a contract. The real catch is the data clause, covered below.

Cloudflare Workers AI advertises “10,000 Neurons/day,” and the word that matters is Neurons, not tokens. A Neuron is Cloudflare’s own compute unit; 10,000 of them buys you several million tokens/day on a cheap model and far fewer on an expensive one. Any list that writes “10,000 tokens/day” for Cloudflare has copied the wrong unit — the official pricing page is clear that it’s Neurons.

Cerebras still gives a flat 1,000,000 tokens/day free, but two things have moved: the free model list rotated to GPT-OSS 120B and Zhipu’s GLM-4.7 (Llama 3.3 70B is no longer on the free tier), and the real constraint is a tight 5 requests/minute and 30,000 tokens/minute ceiling underneath that generous daily total. The rate-limits doc is the source of truth.

Mistral’s La Plateforme free “Experiment” tier still covers the entire lineup including Mistral Large and Codestral — but Mistral has stopped publishing the exact rate numbers on its public docs; you now see your limits only inside the Admin Console. So any specific “500k tokens/min” figure floating around is an estimate, not an official quote. It needs phone/SMS verification, not a card.

I’m keeping Rung 1 deliberately compact here because it has its own dedicated deep-dive. For the per-provider limits, the genuinely-free-on-the-price-sheet models like GLM-4.7 Flash, and the full “what is NOT a free tier” list, read the companion post:

The single best free tier for capable models in 2026 — 1M-token context, no card. Full breakdown of every free API tier, the rate limits, and the data catch lives in Best LLM APIs with a free tier.

Rung 2 — Trial credits (one-time, expiring)

A trial credit is a balance, not a tier: a fixed amount that runs out and does not renew. Useful for evaluating one specific paid platform, useless as ongoing infrastructure. The honest 2026 state here is thinner than most lists claim:

Two things to flag precisely. First, Anthropic has no perpetual free API tier — that part is documented: usage is prepaid, and even purchased credits expire one year from purchase. The specific “$5, expires in 14 days” trial that new accounts see is widely reported and consistent, but Anthropic does not print those numbers on a public page — they surface inside the Console claim flow. So I’ll call it “widely reported” rather than dress it up as an official figure.

Second, Together AI is the one to correct hardest: its docs state plainly that “access to the Together platform requires a minimum $5 credit purchase.” There is no new-account free credit, despite what several aggregator sites claim. If you see “$25 free on Together,” it’s stale or confused with their separate Startup Accelerator. SambaNova is the opposite — its $5/30-day credit sits on top of a permanent rate-limited free tier, which is why it also appears in Rung 1.

Rung 3 — Student deals (and what just died)

If you have a verified school email, this rung is the highest value-per-effort — but it’s also the one where stale lists do the most damage, because student offers change constantly and are fiercely region-specific. Here’s the verified June 2026 state:

The corrections here matter more than the table:

• There is no free ChatGPT Plus for students. That promo ended on 2025-05-31. What’s live is $100 in Codex credits — usable in OpenAI’s coding agent, for verified students residing in the US or Canada. Real, but narrow, and not a Plus subscription.
• GitHub Copilot’s free student plan is paused. The Student Developer Pack itself is alive and still bundles a $100 Azure credit and 100+ partner perks, but the pack page currently states that new Copilot plan sign-ups are “temporarily paused.” And Perplexity Pro, which older lists promise, is not in the current official pack listing — don’t count on it.
• Cursor’s free-year student offer is dead. Cursor discontinued new sign-ups on 2026-06-25 — one day before this post was verified. Existing redeemers keep their rate; new students are pointed to campus event credits or a grad/researcher request form.

Azure for Students is the cleanest of the four: a genuine $100 credit, 12 months, no credit card, renewable while you’re still enrolled, and the credit works against Azure OpenAI (which is how you’d reach GPT-class models with it). The only gate is an accredited, degree-granting institution and a school email.

Rung 4 — Startup model credits (you have a company)

Now the gate changes from “an email” to “a registered company building with AI.” In exchange, the numbers jump. This rung is the AI-specific programs — model and inference credits from the labs and inference providers themselves:

The honesty notes:

• Claude for Startups moved domains — the old anthropic.com/startups now redirects to claude.com/programs/startups. And the official page deliberately does not publish a dollar amount; it says only “apply for free credits and priority rate limits.” The “$1k–$100k” range you’ll see quoted is third-party, not Anthropic’s number. Eligibility for the credits: equity funding from an institutional investor, founded within the last four years, no prior Anthropic startup credits. VC backing is not mandatory to join — only to maximize benefits.
• OpenAI Grove is not an always-on grant. It’s a selective ~5-week, in-person San Francisco cohort (about 15 founders) with up to $50k in API credits as one perk. Its most recent cohort ran January–February 2026, and applications are currently closed — it opens periodically, so it’s a “watch the page” program, not an apply-anytime one.
• Perplexity for Startups is the most gated on this list: the $5k Sonar credits and 6-month Enterprise Pro are real, but you must have raised under $20M, be under 5 years old, and be associated with one of Perplexity’s named Startup Partners. No partner, no application.
• Fireworks AI publishes a startup program with “Build Credits” and engineer access but states no dollar amount anywhere official. The “$5k” and “$10k” figures on aggregator sites contradict each other and aren’t on a Fireworks page — so this post won’t invent one.

For the OSS maintainers reading this: OpenAI’s Codex for open source gives 6 months of ChatGPT Pro (the ~$1,200 figure is just that subscription slice) plus conditional Codex Security access and variable API credits from a $1M fund, if you maintain a widely-used public project. The eligibility is generous — even the official page says to apply and explain your case if your project doesn’t obviously fit.

Rung 5 — Startup cloud grants (the six-figure rung)

The top rung is general cloud infrastructure credit, stage-gated. These aren’t AI-model-specific — they’re compute, storage, and managed services you can point at any AI workload (and on Azure/Google, at the hosted OpenAI and Gemini models directly). The ceilings reach $350k, but the gates get serious:

The corrections that keep this rung honest:

• Google for Startups publishes its tiers cleanly: Start up to $2,000, Scale up to $200k over two years, AI-First up to $350,000. The widely-repeated “$350k over 2 years” is imprecise — $350k is specifically the AI-First ceiling; the standard Scale path tops out around $200k. And you don’t strictly need a VC — eligibility is by company age and funding stage, not partner association.
• AWS Activate’s real tiers: Founders up to $5,000, Portfolio up to $200,000, plus an invite-only AI tier at $200k+. The “$100k” you’ll see elsewhere is outdated; the bootstrapped path (Founders) maxes at $5k, and the big numbers require an Activate provider (accelerator/VC) organization ID.
• NVIDIA Inception is the most mislabeled. It’s free to join and unlocks partner credits (up to $100k via AWS, up to $150k via Nebius) plus roughly 30% off DGX Cloud and $10k in training credits. The common claim “up to $100k DGX Cloud” is wrong on two counts: DGX Cloud is offered as a discount, not a credit, and the $100k is the partner (AWS) credit. The partner amounts aren’t on NVIDIA’s own page and aren’t guaranteed.
• Alibaba Cloud AI Catalyst is the one program here whose status is genuinely in question. The headline — up to 2 billion Model Studio tokens and up to $120k in tiered credits (Launcher $1k → Diamond $59k) — is real, but the official terms state the promotion period ran from 2025-04-01 to 2026-03-31. That window has closed, even though the apply page is still live and aggregators still list it as active. Treat it as pending-renewal and verify before you rely on it.

A note on the two non-LLM entries: Supabase (6 months of the Team plan free, no VC referral needed) and Qdrant (20% off Cloud for a year) aren’t model credits — they’re the database and vector-store layers an AI product sits on. They earn a place here because “free AI compute” in practice also means the infrastructure around the model, and both are easy early wins for a startup with a real website and a company-domain email.

The catches nobody puts in the headline

Across all five rungs, the same handful of catches recur. Knowing them is the difference between “free” and “free until it bites”:

1. Your data is often the price. On free API tiers, assume your inputs train the model unless told otherwise. Google’s free tier says so explicitly; Mistral trains on free consumer-product data unless you opt out (though its API carries a no-training guarantee); OpenRouter’s free models inherit each upstream provider’s policy. This is the catch that should keep proprietary code and customer records off free tiers entirely.

2. The published amount is often a guess. Rungs 4 and 5 hide their real numbers behind an application. When you see a precise “$1k–$100k” for Claude for Startups or a per-stage table for Microsoft, check whether it traces to the official page — frequently it doesn’t. The ceilings are real; the specific figure you’ll be offered is not knowable until you apply.

3. Eligibility is a wall, not a suggestion. Student offers want a verified school email in the right country. Startup programs want a company-domain email, a real website, a company under a certain age, and — for the big tiers — institutional funding or a named partner. A personal Gmail address is the single most common rejection reason on the startup rungs.

4. The fine print changes the math. DigitalOcean’s “free GPU” is three months, after which it’s a discounted $1.90/GPU-hour — and its core credits don’t even cover GPU Droplets. Cerebras’s “1M tokens/day” sits under a 5-requests-per-minute ceiling. Trial credits expire (Anthropic ~14 days, SambaNova 30 days, most cloud credits 12 months). Read the second sentence, not just the headline.

How to pick — by who you are

The fastest path to the right rung is to start from your own situation, not from the list:

1. Hobbyist or learning? Stay on Rung 1. Google AI Studio for capable models, Cerebras or Groq for volume and speed, Cloudflare or GitHub Models if you’re already in those ecosystems. No card, no application, no expiry.
2. Evaluating one paid platform? Rung 2 — grab the trial credit (Anthropic ~$5, SambaNova $5) to kick the tires, then move to a paid plan. Don’t build anything durable on it.
3. A student? Rung 3 is the best value-per-effort: $100 in Azure credit (covers Azure OpenAI) plus the GitHub Student Pack, and $100 in Codex credits if you’re in the US or Canada. Skip anything promising free ChatGPT Plus or free Cursor — both are gone.
4. A solo or bootstrapped founder? The entry tiers of Rung 5 are open to you without VC backing: AWS Founders (up to $5k), Google Start (up to $2k), Supabase, and Claude for Startups. Plus Baseten if you’re AI-core.
5. Funded, with an accelerator or institutional investor? The whole top of Rung 5 unlocks — Google AI-First (up to $350k), AWS Portfolio (up to $200k), Microsoft (up to $150k), Perplexity, and IBM Premium. Apply to several; they don’t conflict.
6. An open-source maintainer? OpenAI’s Codex for OSS is built for you — 6 months of ChatGPT Pro plus credits, with deliberately flexible eligibility.

And the universal rule for every rung: the moment a free tier’s rate limits, data policy, or eligibility stop fitting — when you’ve validated the idea and you’re putting real traffic or real customer data through it — graduate to paid. The paid floor is lower than most people think; the cheapest paid API by use case shows where to land, and tuning reasoning effort cuts the paid bill further once you’re there.

Every figure in this post was verified on 2026-06-26 against each provider’s own page. Free-tier limits, credit amounts, and program eligibility change often — these are the sources, and when in doubt, the provider’s live page wins:

Free tiers: Google AI · Google pricing/data clause · Groq · Cerebras · Mistral · OpenRouter · Cloudflare · GitHub Models · Cohere · SambaNova · NVIDIA. Trials & students: Anthropic · Together AI · OpenAI Codex for Students · Azure for Students · GitHub Student Pack · Cursor. Startup programs: Claude for Startups · OpenAI Grove · Codex for OSS · Baseten · Perplexity · Fireworks · Google for Startups · AWS Activate · Microsoft for Startups · NVIDIA Inception · IBM · DigitalOcean Hatch · Supabase · Qdrant · Alibaba Cloud. See our methodology for how we re-verify.