Overview

OP Inference puts many models behind one API and one key. Text requests use the OpenAI chat completions format, so an OpenAI SDK works by changing its base URL. Media requests (image, video, speech, transcription and music) are jobs: you submit one, then poll it or get a webhook when it finishes.

What the API promisesCopy Anchor Link

These are rules the gateway enforces on every request, not settings you have to remember.

• You get the model you asked for. If an endpoint fails, the request may move to another endpoint serving the same model. It never moves to a different model unless you list one in models.
• No switch after the first token. A request can fail over only before any content reaches you. Once the reply has started, a failure ends the stream with an error frame instead of splicing in another endpoint's answer. See Streaming.
• An empty reply is a failure. A 200 that carries no output is treated as an error and can fail over.
• Retention is per request. Every request runs under a retention class: zdr, short or standard. A request can make its class stricter than your key's, never looser, and the class is enforced everywhere the request is stored or logged. See Data retention.
• One key for every modality. Text, image, video, speech, transcription and music share one key, one balance and one bill.

Base URLCopy Anchor Link

https://api.opinference.com/v1

All requests authenticate with Authorization: Bearer <key>. See Authentication and API keys.