OpenAI Compatibility

​

Unchanged

• Request shape. messages, temperature, top_p, max_tokens, max_completion_tokens, stream, tools, tool_choice, user.
• Response shape. A chat.completion with choices, message, finish_reason, usage.
• Streaming. stream=True / stream: true, Server-Sent Event chunks.
• Tool calling. The model returns tool_calls; your code runs them and returns a tool message.
• Prompt caching. cache_control on content parts, on routes that support it. See Prompt Caching.
• Reasoning. reasoning_details in the response, reasoning_tokens in usage.
• Model listing.

​

Differences

• system_fingerprint is always null.
• Exactly one choice; n > 1 is rejected.
• The error envelope is {"error": {"message", "code", "param"}} with no type field.
• The privacy level travels with the model (:confidential); there is no privacy parameter. Encrypted adds client-side encryption and request headers on top.
• /v1/chat/completions drops unknown fields at gateway-owned objects. Controls the model cannot honor return 400.
• Rate-limit headers: X-RateLimit-Limit, X-RateLimit-Remaining, X-RateLimit-Reset, and Retry-After on 429. See Rate Limits.
• OpenAI model names are not Nito IDs. See model mapping.

​

Other Surfaces

• Responses API (/v1/responses): a stateless shim. store: true and previous_response_id are rejected. See the Responses reference.
• Files API (/v1/files): not served, 404. Send documents inline. See Multimodal.

​

Rejected Inputs

• Audio and video input.
• file_id content references. Supply file content inline.
• response_format (json_object, json_schema) on models that do not advertise json_mode or structured_outputs.

​

Related