Use the authenticated catalog
Call GET /v1/models with the same key that will make the completion request. Use the returned id exactly. Model aliases are owned by Accountable and do not necessarily match an upstream provider ID.
curl "$GATEWAY_BASE_URL/v1/models" \
-H "Authorization: Bearer $GATEWAY_API_KEY"Public library versus callable models
The public model library includes planned models and readiness information. It is not an API allowlist. A model shown there can still be awaiting provider access, a connected credential, retention review, or commercial approval.
| Status | Meaning |
|---|---|
| Internal test verified | A controlled synthetic request has worked; customer or PHI availability does not follow automatically. |
| Access requested | Provider access is still pending. |
| Integration pending | A direct provider test worked; the Gateway route and healthcare requirements are still pending. |
| Connection pending | Provider setup is underway; a Gateway call is not yet verified. |
| Under review | The candidate is being evaluated. |
Why a listed model can still fail
Discovery omits models marked unavailable and models outside your key’s allowlist. Other checks, such as credits, account activation, data permissions, and current healthcare policy, happen when you submit a request. A discovery response is not a certification of PHI eligibility.
Switch without changing the API shape
Change only the model value after checking its permissions, pricing, and input/output limits. Preserve the request schema. A new model can behave differently even when the transport format stays the same; evaluate it on your own representative synthetic tasks. There is no automatic fallback to a different model.
OpenAI in private staging
21 direct OpenAI models have passed synthetic verification. The model catalog shows their provider prices and verified streaming, structured-output and tool capabilities.
Displayed model rates exclude the platform fee; usage billing adds 15%. Each route uses the isolated Gateway organization and Standard processing, with model-specific request parameters and exact approved response versions. The Gateway rejects unexpected cached usage rather than silently applying an incorrect charge.
Only synthetic text is enabled. BAA acceptance does not establish account-specific ZDR, regional processing approval or customer distribution permission. Responses-only models, image/audio endpoints and embeddings are not yet exposed. Retired OpenAI models are excluded even when a discovery endpoint still lists them.