A rate limit that counts the tokens you asked for, not the ones you got

A developer discovered that Groq's free tier rate limit counts the declared max_tokens rather than actual generated tokens, causing requests with small prompts but large max_tokens values to be rejected with a 413 error. Testing across 14 models revealed that four models failed identical requests when max_tokens was set to 8192, while passing with a lower value, and that the limit is a rolling window shared across models. The developer advises setting max_tokens to actual needs, such as 1500, to avoid failures in agent applications.

Spent a while assuming a 413 meant my prompt was too long. It wasn't. The prompt was 20 tokens. Groq's free tier caps you at 8,000 tokens per minute, and that budget is charged against the max tokens you declare , not the number the model actually generates. So this fails: php prompt: 20 tokens, max tokens: 8192 - 413 "on tokens per minute TPM : Limit 8000, Requested 8271" Nothing was generated. The request was rejected on the ceiling I asked for. The same model with max tokens: 16 and a 4,078-token prompt returns 200 without complaint. Prompt size was never the problem. I swept the 14 models Groq advertises. Four of them answered a minimal request in ~300ms and then returned 413 on the identical request with max tokens: 8192 : php openai/gpt-oss-120b 478ms - 413 Requested 8271 qwen/qwen3.8-27b 324ms - 413 Requested 8212 qwen/qwen3.6-27b 330ms - 413 Requested 8210 openai/gpt-oss-safeguard-20b 308ms - 413 Requested 8271 Two more passed the same request, which is the second thing worth knowing: it is a rolling window shared across models, not a per-model constant. A model that passes one probe can 413 a minute later — one of mine came back Limit 8000, Used 4373, Requested 6278 . Never conclude a model is exempt from a single passing call. The practical consequence is for agents. Tool schemas already eat the prompt, and most harnesses declare a generous max tokens by default. Set it to what you actually need — 1,500 works — or every call dies before the model sees it. Worth checking whether your provider bills the declared ceiling or the real output. Alibaba's docs, for contrast, say plainly that their TPM "includes input and output tokens."