Utmost
Utmost is an API for open-weight language models. It serves each model in two variants: the standard checkpoint, and an abliterated checkpoint with the refusal behaviour removed. It is built for research and development teams whose work makes a typical assistant refuse: security research, red-teaming, safety evaluation, and medical, legal and policy work.
The API is compatible with the OpenAI Chat Completions and Responses APIs and the Anthropic Messages API. Existing SDKs and tools work after a base-URL change.
Who it is for
- Standard models are open to any organization with an account (Developer tier).
- Abliterated models need a Verified organization: a confirmed legal identity and a signed acceptable use policy. See Access tiers.
- Bring-your-own-model needs the Research tier.
Two hard limits apply on every tier and every model: child sexual abuse material, and serious uplift for biological, chemical, nuclear or radiological weapons. See Safety.
What is different
Eval cards before traffic. A model receives no traffic until its eval card is published. The card shows the refusal rate on a named prompt set, capability scores for the abliterated checkpoint next to its base model, the method used, and the checkpoint hashes. See Models.
Routing. Send model: "auto" and the router picks the cheapest tier likely to handle the request. If a standard model refuses a non-streaming request, the router retries once on the abliterated variant, and the refused attempt is not billed. Every response says which model ran and what it cost. See Routing.
Bring your own model. Submit a Hugging Face checkpoint. We check its license, abliterate it, run the eval harness, and show you the report. Once you approve it, it is served privately to your organization. See Bring your own model.
Zero retention by default. We store request metadata (tokens, cost, latency, model), not prompts or outputs. Storage is opt-in. See Data retention.
Start here
- Quickstart: an API key and a first call in a few minutes.
- Authentication: keys, projects, and scoping.
- Chat completions: the main endpoint.
- Routing: how the
autoaliases choose a model. - Errors: every error code, and which ones to retry.
API at a glance
| Endpoint | Purpose |
|---|---|
POST /v1/chat/completions |
OpenAI-compatible chat completions |
POST /v1/responses |
OpenAI Responses API |
POST /v1/messages |
Anthropic-compatible Messages API |
GET /v1/models |
Models, prices and capabilities (no key needed) |
POST /v1/batches |
Asynchronous requests at 50% off |
POST /v1/workflows/security-review |
Multi-pass security review |
POST /v1/byo/models |
Bring your own model |
GET /v1/credits |
Balance, spend and caps |
The machine-readable OpenAPI 3.1 description covers every endpoint above.