Connection recipes

Start with the API family, then choose the client. OpenAI Chat Completions, Responses and Anthropic Messages use different payloads even when the credential is the same. A hostname change alone does not translate one protocol into another.

Client setup

Acceptance before rollout

Test an enabled model in Playground, verify streaming if the application expects it, then configure the client using a dedicated key. Enable only capabilities actually advertised or tested for the selected model. Tool-calling, vision and reasoning are not implied by a familiar model name.

Model Verification uses a ready behavioral reference selected by canonical model. It sends 32 saved questions in one browser request; neither cutoff lookup nor question generation occurs. An incomplete reference blocks the test instead of charging for preparation. The reference pins protocol and request parameters, and the external endpoint key stays browser-local. The judge receives the saved observations and current answers, then alone assesses semantic correctness and correspondence in both directions, including unexpected correct answers. A formatting irregularity is not an automatic semantic penalty. Evaluation uses an eligible company key under existing role, quota and wallet rules; reference preparation is not charged as part of your verification request. No behavioral result proves exact provider identity. An application with a fixed origin may need the optional local endpoint override. Prefer native endpoint configuration whenever available.

Unsupported API paths now return 404 endpoint_not_supported with the requested path; a wrong method on a supported path returns 405 method_not_allowed and Allow. Native /v1/balance is unchanged; /v1/user/balance is not implemented. Share the returned request ID, never your API key, when reporting these errors.

Request builder choices for team keys

Select the intended request type before composing a request. The workspace narrows model, API-family, endpoint and response-mode choices using model metadata and checks for that particular protocol. For Claude, the native Anthropic Messages endpoint is preferred when the model and request type permit it. A negative check for one endpoint cannot be overridden by a successful result for another. This selection does not grant extra company/project permissions or establish current provider availability.

The explanation beside Send identifies a missing input, unavailable compatible combination, invalid token limit or a submission already in progress. Raw JSON is a separate explicit advanced mode. The builder can compose Vision and function calls for a selected Chat Completions endpoint, but does not advertise that endpoint for PDFs. Team access and live API admission remain authoritative; Model Verification keeps its own pinned protocol.

Dedicated Images tab

Company members with executable project keys can use the separate Images workspace for reviewed-route generation/editing, quote confirmation, saved state and private files. This does not expose a colleague's secret, grant company-wide file access or turn Model Verification into a background task.

Integration scope

Company tool instructions do not certify every current editor release or its image support. Audit outbound host, actor/project credential and retry behavior for the installed client. A text-agent configuration does not implement the company Images quote/state/private-file protocol automatically; follow its explicit HTTP contract and do not retry an ambiguous paid POST.

For recordings, use Playground Audio and the transcription API. Audio is separate from text chat and Images.

Video workspace

Playground → Video provides text-to-video quotes, progress, saved MP4 playback, private downloads and Raw/cURL. Use an available model/route for the selected key; another paid operation is not a retry of an unknown outcome.