OpenAI compatibility
What works when you point an OpenAI SDK at BlazeInfer, and the differences to watch for.
BlazeInfer supports the OpenAI Chat Completions API. For most applications, using it means changing the base URL, the key and the model ID.
Endpoint matrix
| OpenAI call | Endpoint | BlazeInfer status |
|---|---|---|
| Chat completions | POST /v1/chat/completions | Supported — the main path, including streaming, tool calling and structured outputs |
| List models | GET /v1/models | Supported — returns full model IDs in the OpenAI list shape |
| Responses | POST /v1/responses | Enabled per model. See the model listing in the console |
| Messages (Anthropic) | POST /v1/messages | Enabled per model. See the model listing in the console |
Known differences
These are the behaviours that differ from OpenAI and that most often surprise OpenAI-compatible tools:
- Streaming usage. BlazeInfer adds a final usage chunk when the request has no
stream_options. If you sendstream_options, setinclude_usage: true. Handle that extra chunk with an emptychoiceslist. - Aliases. BlazeInfer accepts short model aliases, which OpenAI does not have. Use full IDs for production.
- Reasoning field name can differ between models. Reasoning models may return
reasoning in
reasoning_contentorreasoning. Treat both as possible until you have tested your model. See Reasoning models.
Capabilities vary by model
Tool calling, reasoning field names, structured outputs and vision all depend on the model. Check the console and the guides before you rely on them.