API REFERENCE
Request parameters
Sampling, output limits, model options, and validation.
Shared parameters
| Parameter | Accepted values | Notes |
|---|---|---|
| model | Supported ID | Bare OpenAI IDs and openai/ prefixed IDs |
| temperature | 0–2 | Omitted values use upstream defaults; reasoning models may reject it |
| top_p | 0–1 | Tune this or temperature, usually not both |
| stream | true / false | SSE is available on text endpoints |
| store | false / omitted | No server-side response retention; true returns 400 |
| service_tier | default / auto | Explicit flex, priority, and other requested tiers are rejected |
| parallel_tool_calls | boolean | Model-dependent; relevant when using tools |
| metadata | string-keyed object | Upstream metadata; not a substitute for local access control |
| user | string | Application-level end-user identifier; do not include secrets |
Only specified sampling settings are forwarded. Valid ranges do not guarantee that every model supports every option. Model-specific validation errors come back in the provider’s error envelope.
Chat Completions parameters
| Parameter | Values / shape | Notes |
|---|---|---|
| messages | Non-empty array | Full conversation including tool results |
| max_completion_tokens / max_tokens | Positive integer | Up to the model’s output limit, with an absolute 200,000 ceiling |
| response_format | text / json_object / json_schema | Schema is nested under json_schema |
| reasoning_effort | Model-specific effort string | Examples: low, medium, high; availability varies |
| frequency_penalty / presence_penalty | −2 to 2 | May not be supported by reasoning models |
| seed | Integer | Best-effort reproducibility, not a guarantee |
| logprobs / top_logprobs | Boolean / integer 0–20 | top_logprobs requires logprobs on compatible models |
| logit_bias | Token-ID → bias map | Native OpenAI behavior; typical bias range −100 to 100 |
| stop | String or array | Model-dependent stop sequences |
| prediction | type=content, content=text or text parts | See Predicted Outputs |
| n | 1 only | Multiple candidate completions are not supported |
Responses parameters
| Parameter | Values / shape | Notes |
|---|---|---|
| input | String or non-empty array | Required even with instructions |
| instructions | String | Re-send on each independent request |
| max_output_tokens | Positive integer | Includes reasoning and visible output |
| reasoning | Object | effort, summary and compatible model-specific options |
| text.format | text / json_object / json_schema | name, schema and strict sit directly in format |
| text.verbosity | Model-specific verbosity | Usually low, medium, or high |
| include | Allowed string array | reasoning.encrypted_content, message.output_text.logprobs |
| truncation | disabled / auto | Upstream context handling |
| tools | Client-executed definitions | No hosted provider execution |
| tool_choice | none / auto / required / specific tool | Shape depends on the tool type |
Explicitly unsupported options
provider, models, route, plugins, transforms, and upstream-body debug are router-specific and return 400. Non-OpenAI sampling options such as top_k, min_p, top_a, and repetition_penalty are also rejected.
Stored response/item references, provider file IDs, hosted tools, and background generation are unavailable. Do not send them as a fallback: a supported schema with an unsupported feature is still an error. See the compatibility table.