“Who gives the most for free?” has no single answer because providers meter different things — requests, tokens, calls or normalized compute units. Here’s the verified landscape, then the winner per category.
The free tiers, side by side
| Provider | Free allowance (verified) | Card | Metered in |
|---|---|---|---|
| Google Gemini API | Selected models; per-model limits shown only in AI Studio | Not stated | Requests |
| Groq | Published per model (e.g. gpt-oss-120b: 30/min, 1,000/day) | Not stated | Both |
| OpenRouter | 50 ‘:free’ requests/day, 20/min (1,000/day after buying $10) | No card | Requests |
| Cohere | 1,000 calls/month (Chat 20/min, Rerank 10/min) | Not stated | Calls |
| Cloudflare Workers AI | 10,000 Neurons/day, resets 00:00 UTC (8 models paid-only) | No card | Neurons |
| Mistral | Included monthly usage; Mistral publishes no figure | No card | Not published |
| OpenAI | No general free API tier | — | — |
| Anthropic (first-party API) | Effectively prepaid; no confirmed signup credit | — | — |
| Cerebras | No longer a free tier — $5 trial, expires in 30 days | Required | — |
| Hugging Face | Ended 7 Oct 2026 — free accounts now get no monthly credit | — | — |
“No card” means the provider says so in its own words; “Not stated” means it does not say either way.
Most complete first-party tier: Google Gemini
The Gemini API free tier is the only free tier from a frontier lab: Gemini 3.8 Flash, Gemini 2.5 Pro, Gemma 4 and Gemini Embedding 2 among others, though Google marks some models — Gemini 3.1 Pro Preview, Omni Flash, Veo, Lyria and the Nano Banana image models — as “Not available” on the free tier. Per-model limits depend on your usage tier and are shown in AI Studio rather than published as one table. You must be 18 or over.
This page compares what each tier gives you. For the limits that decide whether a tier survives your traffic — requests per minute, per day, and what each provider meters — see free LLM API rate limits compared.
Fastest tokens: Groq · Widest choice: OpenRouter
Groq’s Free plan publishes real per-model numbers (openai/gpt-oss-120b: 30 requests/min, 1,000 requests/day, 8K tokens/min, 200K tokens/day) on unusually fast hardware. OpenRouter trades volume for variety: only 50 free requests a day (1,000 once you have bought $10 of credits), but across a rotating multi-provider catalogue behind one API.
Specialists: Cohere, Cloudflare, Mistral
Cohere’s trial keys are the free-embeddings pick, within 1,000 calls a month — Check your account’s current Embed limit in the dashboard. Cloudflare Workers AI gives 10,000 Neurons every day on the Free plan, ideal if you’re already at the edge — though eight large third-party models (two Kimi, three GLM and three DeepSeek V4 variants) “require a paid billing method” regardless of the free allowance. Mistral’s Free mode is the default for new accounts and needs no card, but Mistral no longer says how much usage it includes.
Two that left the table
Cerebras stopped being a free tier in September 2026: its docs now say new accounts receive “$5 in free credits after adding a verified payment method”, that the credits “expire 30 days after they’re granted”, and that there is no permanently free tier. It is still a fast trial credit. Hugging Face ended its $0.10 monthly Inference Providers credit for free accounts — its docs have listed none since 7 October 2026. Both are recorded in the change log.
The honest bottom line
For most people the practical answer is a stack, not a single winner: Gemini as the daily driver, Groq for open-model volume and speed, OpenRouter for model variety, Cohere for embeddings. Holding several keys costs nothing — see how to combine free tiers legitimately and the live directory for current numbers.