Free-tier tables all lead with requests per day, and it's the least binding number on the sheet.
Groq is the clean example: 1,000 requests/day sounds generous, but chat models also carry 200,000 tokens/day. Divide: 200 tokens per request if you actually use all 1,000. A realistic exchange — system prompt, context, a few hundred tokens back — is 2,000+ tokens, so the day really ends around 100 requests. The token ceiling binds at 10% of the request ceiling.
And the token cap is the one that isn't in the response headers — you see requests-remaining, not tokens-remaining, so the binding constraint is the invisible one.
The exceptions are worth knowing: qwen3.8-27b on Groq gets 2M TPD (10x its siblings), and some providers cap only requests. That's the comparison that predicts real capacity.
The full table, sorted by the number that matters: https://toolfreebie.com/best-free-ai-apis-2026/
Top comments (0)