Free LLM APIs That Actually Stay Free

Most "free LLM API" articles were written by scraping each other. We checked every provider's own documentation and pricing pages directly on September 16, 2026 — and dropped everything that doesn't survive the two questions that matter: does it reset, and can you keep using it without paying? What's left is shorter than you'd expect, and one popular entry is a ghost.

The tiers that actually keep giving

1. OpenRouter :free models — the best free tier right now

OpenRouter hosts dozens of models with a :free suffix — free variants of DeepSeek, Qwen, Llama, and more, all behind one OpenAI-compatible endpoint.

The limits, straight from OpenRouter's limits documentation (verified today from the page source):

Credits purchased (all time)	Requests/min	Requests/day
Under $10	20	50
$10 or more	20	1,000

Three things worth knowing that most roundups skip:

• The daily cap is tied to lifetime spend, not a subscription. Buy $10 of credits once — even if you spend it — and your free-model allowance jumps from 50 to 1,000 requests a day. That $10 also unlocks the paid models, so it's arguably the cheapest quality-of-life upgrade in the free-tier game.
• A negative balance breaks free models too. If your paid credits go below zero, OpenRouter returns 402 errors even on :free models until you top back up. Free does not mean unconditional.
• Rate limits reset daily in UTC.

For a site that tracks DeepSeek pricing closely (see our cheapest DeepSeek API provider comparison), the :free DeepSeek variants are the zero-cost entry point — just expect the free pool to be slower and busier than paid routing.

2. Cloudflare Workers AI — 10,000 Neurons a day, no card

Workers AI is available on the Workers Free plan with 10,000 Neurons per day at no charge, resetting daily at 00:00 UTC. Neurons are Cloudflare's cross-model compute unit — roughly "how much GPU work your request took" — and on the paid plan they cost $0.011 per 1,000.

The catch, verified on Cloudflare's own pricing page today: a list of the big frontier models is excluded from the free allocation. Kimi K2.6/K2.7-Code, GLM-5.2, GLM-5.3, GLM-5.3-Flash, DeepSeek V4 Flash 0731, and DeepSeek V4 Pro 0813 all require either the Workers Paid plan or prepaid AI Gateway credits. So the free 10k Neurons buy you the smaller catalog — genuinely useful, genuinely capped.

That exclusion list is a quiet but real shift: the exact models we cover in our GLM-5.3-Flash field report are pay-walled behind Cloudflare's cheapest tier.

3. Mistral's Experiment tier — the fat but fuzzy one

Mistral's free "Experiment" plan on La Plateforme historically includes around 1 billion tokens per month with low per-minute limits — no credit card required. We could not reach Mistral's docs from our network today, so treat the 1B figure as the consistent third-party consensus (pricepertoken, cloudzero, others), not a number we verified against the source. If you're about to build on it, check the limits shown in Mistral's console for your workspace — per their own docs, exact RPM is displayed there.

The ones that don't qualify

The defining rule of this list: a one-time signup credit is not a free tier. It's a coupon. It runs out and never comes back:

• Cerebras — self-serve Developer tier comes with a one-time $5 credit, then pay-as-you-go ($0.35/$0.75 per M tokens for GPT-OSS 120B at ~3,000 tokens/s — genuinely fast, genuinely paid). No ongoing free allowance as of today. Their pricing page confirms the tier structure.
• Novita — new accounts get a small one-time signup credit. No standing free tier. What Novita does have is some of the lowest paid prices on DeepSeek's V4.1 Flash line — $0.14 input / $0.28 output per M tokens on DeepSeek V4 Flash as checked today — which matters if you graduate from free tiers to real usage (that's the comparison we run in our Qwen3.8 Flash vs DeepSeek 4.1 Flash breakdown).

The trap: dead free tiers still ranking on Google

GitHub Models is gone. GitHub's own documentation states that as of July 30, 2026, GitHub Models was fully retired — the playground, model catalog, inference API, and BYOK are no longer available to any customer, and GitHub points developers to Azure AI Foundry or Copilot instead.

We're calling this out because GitHub Models still appears in "best free LLM APIs" lists that haven't been touched since spring. If a roundup recommends it without a dated caveat, that's your signal about how stale the rest of the list is.

Groq, Gemini, and the limits we couldn't pin down

Honesty section, because every list should have one.

• Groq's free tier is real — every provider we could reach agrees on roughly 30 requests per minute — but the daily cap is a mess: one regularly-updated comparison (Sept 4, 2026) says 14,400 requests/day, another (May 2026) says 1,000 requests/day with 200K tokens/day. Groq's own rate-limits page was not reachable from our network today, so we won't pick a winner. Two sources disagree; both are cited.
• Gemini's free tier exists on the AI Studio API key path, but Google applies limits per project, not per key, resets daily quotas at midnight Pacific time, and sets wildly different caps per model. Third-party reporting notes Google slashed free limits in December 2025 (Gemini 2.5 Flash reportedly down to ~20 requests/day) while newer Gemini 3 Flash models reportedly get ~1,500 requests/day on the free tier. We couldn't verify either number against Google's docs from here — check the official rate-limits page before committing a workload.

Both of these still belong in the free-tier conversation. They just don't belong with unverified numbers attached.

Summary table

Provider	Ongoing free offer	Verified today?	Reset	Gotchas
OpenRouter :free	50 req/day (1,000 if lifetime spend ≥ $10), 20 req/min	Yes — official docs	Daily (UTC)	Negative balance blocks free models too
Cloudflare Workers AI	10,000 Neurons/day	Yes — official docs	Daily (00:00 UTC)	GLM-5.x / DeepSeek V4 / Kimi K2 models excluded
Mistral Experiment	~1B tokens/month	No — third-party consensus	Monthly	Low RPM; exact limits shown in console
Gemini (AI Studio)	Yes, per-model	No — official page unreachable	Midnight Pacific	Per-project limits; caps vary hugely by model
Groq	Yes, ~30 req/min	Partially — daily cap disputed	—	14,400 vs 1,000 req/day across sources
Cerebras	None — one-time $5 credit	Yes — official pricing	Never	Pay-as-you-go after credit
Novita	None — one-time signup credit	Yes — official pages	Never	Cheap paid tier instead
GitHub Models	Retired July 30, 2026	Yes — official docs	—	Don't build on it

What to actually do with this

If you're prototyping: OpenRouter's :free catalog plus a Cloudflare Workers free account covers most weekend projects at literal zero cost. Once your usage outgrows 50 requests a day, the honest options are the $10-lifetime OpenRouter unlock or moving to a cheap paid provider — and that's where the real pricing work starts, in our DeepSeek API cost calculator and the off-peak pricing explainer (same model, half price, if you can run at night).