The paranoid stack
For prototypes touching anything sensitive: every rung below is confirmed (not "unknown") to require no credit card and to not train on your prompts. That second constraint is stricter than it sounds — most free tiers either train, log, or won't say.
The catalog treats "unknown" as failing this filter. As providers get verified, the set grows:
pnpm mh list --no-card --no-training.
Confirmed today (2026-07 snapshot)
| Provider | Model | Notes |
|---|---|---|
| OVHcloud AI Endpoints | gpt-oss-120b, Meta-Llama-3_3-70B-Instruct, Qwen2.5-VL-72B-Instruct (vision) |
Official policy: "Your data will never be used to train…", zero retention. Anonymous tier ~2 RPM/IP; keyed tier lifts it. |
One confirmed provider is the honest answer right now. Notable near-misses:
- Groq — no card, but training policy not yet confirmed on an official page (
prompt_logging: unknown). Likely fine; verify before trusting. - Google AI Studio — explicitly may train on free-tier prompts outside UK/CH/EEA/EU. Paid tier doesn't. Know which one you're on.
- OpenRouter
:freemodels — upstream free providers "may log prompts for training" per OpenRouter's own docs. - Mistral free tier — requires opting into data sharing. The opposite of this stack.
Wire it
pnpm mh recommend --no-card --no-training --snippet
# OVHcloud anonymous — zero signup, rate-limited; keyed access removes the /IP cap
curl https://oai.endpoints.kepler.ai.cloud.ovh.net/v1/chat/completions \
-H 'content-type: application/json' \
-d '{"model": "gpt-oss-120b", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 64}'
Recheck after each verification pass — this recipe is regenerated from the catalog, and providers move between "unknown" and confirmed as probes and doc-fetches land.