~/providers
Which AI should I use?
Free-tier limits and real per-run costs across all 9 providers — plus what your own history says about each one.
No invented scores. There is no ground truth for ad copy, so this page never claims one model is “better”. Every figure is either a published fact (price, free cap, vision) or measured from your own generations. The most useful signal is retry rate: how often a provider’s first reply failed validation and had to be re-asked — that costs you double and is genuinely observable.
| provider | free tier | free req/day | ~cost / run | vision | your runs | your spend | retry rate |
|---|---|---|---|---|---|---|---|
OpenRouter google/gemma-4-31b-it:free | Yes · no card Applies to :free models only. Paid models draw on your OpenRouter credit. | 50 | free | — | — | — | |
Groq openai/gpt-oss-120b | Yes · no card | 14,400 | $0.0023 | — | — | — | |
Cerebras gpt-oss-120b | Yes · no card Daily token cap is not published — expect throttling on heavy use. | 30/min | $0.0023 | — | — | — | |
Google Gemini gemini-3.6-flash | Yes · no card Free limits are PER MODEL. gemini-3.6-flash allows only ~20 requests/day free; the Flash-Lite models are far more generous. | 20 | $0.03 | — | — | — | |
DeepSeek deepseek-v4-flash | Paid only | — | $0.0015 | — | — | — | |
Together AI openai/gpt-oss-120b | Paid only Signup credit only, not an ongoing free tier. | — | $0.0023 | — | — | — | |
Mistral mistral-medium-3-5-26-04 | Paid only Free signup credit, then pay-as-you-go. | — | $0.03 | — | — | — | |
OpenAI gpt-5.6-terra | Paid only | — | $0.04 | — | — | — | |
Anthropic Claude claude-sonnet-5 | Paid only | — | $0.05 | — | — | — |
“~cost / run” uses 8,000 tokens per generation (default estimate; run a few generations to personalise) at a 70/30 input:output split, priced against each provider’s default model. Free-tier caps are published figures and change often — verify with each provider before relying on them.
starting from zero? read this
- Most free requests/day: Groq — 14,400/day, no card. The practical choice for high-volume work at $0.
- Largest free models: OpenRouter’s
:freetier and Gemini both expose frontier-class models at $0 — but Gemini’s cap is per model and as low as ~20 requests/day, and OpenRouter’s is 50/day. Fine for drafting, not for volume. - Need image input? Only Claude, GPT, Gemini and OpenRouter accept screenshots. Groq and Cerebras are text-only.
- Best for strict JSON: watch the retry-rate column above once you have a few runs on each — that is real evidence, unlike a vendor claim.