Every “best free LLM API” list ranks tiers by how big they sound. That is the wrong axis. A free tier fails because of the shape of your traffic — a burst that exceeds a per-minute cap, a long tail that exhausts a daily one, a single enormous prompt that blows a token ceiling — and those are three different limits that bind at three different moments.
So here is the same question asked properly: what does each provider actually meter, and what are the published numbers? Everything below was verified against the current offer records and their official sources on 21 October 2026, and each row links to the offer page carrying its primary source.
The published free limits, side by side
| Provider | Requests/min | Requests/day | Tokens | Meters | Card |
|---|---|---|---|---|---|
| Groq | 30 | 1,000 | 8K/min · 200K/day | Requests + tokens | Not stated |
| NVIDIA NIM | Up to 40 for most models | Not stated here | Model-specific | Requests | Not stated |
| OpenRouter | 20 | 50 (1,000 after $10 spend) | — | Requests | No card |
| Cohere | 20 (Chat) | — | — | Calls/month (1,000) | Not stated |
| Cloudflare Workers AI | — | — | — | 10,000 Neurons/day | No card |
| Gemini free tier | Not published | Not published | Not published | Requests | Not stated |
“No card” means the provider says so in its own words; “Not stated” means it does not say either way.
Groq’s figures are for openai/gpt-oss-120b; openai/gpt-oss-20b and
qwen/qwen3.8-27b share 30 requests/min, 1,000 requests/day, 8,000 tokens/min and
200,000 tokens/day, while openai/gpt-oss-safeguard-20b is tighter at 3
requests/min. Groq’s agentic groq/compound systems, which used to appear here,
were decommissioned on 21 September 2026.
One entry is in the table to be honest about what it is, not because it is comparable: Cloudflare meters Neurons, a normalised compute unit, so your real capacity depends entirely on which model you call. Hugging Face’s $0.10 monthly balance used to sit beside it; its docs have listed no free-account credit since 7 October 2026.
And Google publishes no numbers. The Gemini free tier is genuinely free, but its per-model limits vary by usage tier and are shown inside AI Studio rather than documented publicly. We could publish a number we read off one account, but it would not be a fact about your account, so we do not.
Match the meter to the workload
Request-heavy and interactive — a chat bot, an autocomplete, anything a person waits for. You are bound by requests per minute first. Groq at 30/min has a clearly documented worked example, and it runs on hardware fast enough that latency is a feature rather than a tax. The number to design against is the daily cap of 1,000, not the per-minute one.
Bursty and unpredictable — anything with public traffic. No free tier survives this, because they all cap per minute and a burst fails rather than queues. Hold keys at two or three providers and fall back between them; their quotas are independent. Our rate-limit guide covers the retry discipline that stops a burst becoming a burned day.
Few, large calls on a schedule — a nightly summariser, a document pass, an eval run. Per-minute caps barely matter; the daily ceiling does. Cloudflare Workers AI resets 10,000 Neurons at 00:00 UTC every day; this is a reset, not a guarantee that the programme will continue forever.
Embeddings, search and reranking — a different shape again, because you send many small inputs. Cohere is the specialist: 1,000 API calls a month on a trial key, with per-endpoint ceilings underneath (Chat 20/min, Rerank 10/min, Tokenize 100/min). Check the current trial Embed limit in your dashboard. Note the month-long meter — there is no daily reset to wait for, so a bad afternoon costs you the rest of the month.
Trying many models — OpenRouter is the
variety buffet: one key, a rotating catalogue of :free models, 20 requests a
minute. The catch is volume: 50 requests a day until you have bought $10 of
credits at some point in the account’s life, after which it is 1,000/day
permanently. It is the best prototyping tier here and the worst production one.
Three traps the numbers do not show
The tightest limit wins, and it is rarely the one you were watching. Groq’s 30 requests/min and 8,000 tokens/min are simultaneous ceilings: thirty requests carrying 300 tokens each will hit the token limit first. Check the shape of every limit, not the size of the headline one.
Failed requests can still cost you. OpenRouter states that “Failed attempts count toward that daily cap”. A retry loop without backoff converts one bad request into an exhausted day.
A free tier can stop being free. This is not hypothetical — it happened twice in two months. Cerebras was in this comparison until September 2026 on the strength of 1M tokens/day per model; its documentation now states there is no permanently free tier, that the $5 in credits arrive only “after adding a verified payment method”, and that they expire 30 days later. Then, on 7 October 2026, Hugging Face’s docs dropped the free-account monthly credit altogether. The change log records what moved on each pass; check the last-verified date on any table before you build against it, including this one.
The verdict
There is no single winner, because there is no single meter. If you want one default, take the Gemini free tier for breadth and Groq for anything where requests per minute is the binding constraint — then add OpenRouter for variety and Cloudflare for scheduled volume. Holding all four keys costs nothing, and their quotas are independent, which is the only real protection against any one of them changing.
For the full picture, see free AI API tiers, the side-by-side of what each tier gives you, and how to combine free tiers legitimately.