Request parameters

The gateway forwards to the model only the parameters on its endpoint's list. Anything outside the list is silently stripped: the request runs as if it were absent. So a parameter you are used to in one format may quietly disappear in another: for example, only /v1/chat/completions accepts seed, and max_tokens is called max_output_tokens on /v1/responses.

Forwarded to the model#

The gateway validates only the model field; the values of other parameters are validated by the model's provider. If it rejects a value, the response is a 400 with the unsupported_value or invalid_value code (see Errors).

/v1/chat/completions#

model, messages, stream, temperature, top_p, max_tokens, max_completion_tokens, stop, seed, n, presence_penalty, frequency_penalty, logit_bias, logprobs, top_logprobs, response_format, tools, tool_choice, parallel_tool_calls, functions, function_call, prediction, reasoning_effort, verbosity, prompt_cache_retention, top_k, min_p, repetition_penalty, thinking, enable_thinking, thinking_budget, reasoning

/v1/messages#

model, messages, system, max_tokens, stop_sequences, stream, temperature, top_p, top_k, tools, tool_choice, thinking, cache_control, output_config (only effort and format; the legacy output_format moves to output_config.format)

/v1/responses#

model, input, instructions, stream, temperature, top_p, top_logprobs, max_output_tokens, max_tool_calls, tools, tool_choice, parallel_tool_calls, reasoning, text, truncation, include, prompt_cache_retention

/v1/images/generations#

model, prompt, n (clamped to 1–10), size, aspect_ratio, resolution, quality, background, image, images, output_format, output_compression, response_format

/v1/videos#

model, prompt, seconds, duration, duration_seconds, size, aspect_ratio, resolution, image, images, video, input_reference, negative_prompt, seed

Special cases#

• stream_options is stripped, but "include_usage": true on /v1/chat/completions is honored: usage arrives in the last stream chunk.
• user, safety_identifier, prompt_cache_key and metadata are not passed to the provider — the gateway sets its own identifiers. Different prompt_cache_key values still map to different cache partitions.
• On /v1/responses store is always false: conversations are not stored at the provider.
• Parameters that change the provider's pricing tier or routing are stripped without an error: service_tier, inference_geo, speed, models, provider, route, transforms, plugins, as well as web search and code execution (web_search_options, mcp_servers, container and the like).

Tools#

Only tools that your code executes, or that cost nothing but tokens, stay in tools. The provider's built-in tools (web search, code execution, file search, MCP) are silently removed from the list.

If the list ends up empty, tools, tool_choice and parallel_tool_calls are removed and the model answers without tools. But if tool_choice requires a tool that is no longer there (required, a specific tool, any), the request is rejected — see below.

Rejected with a 400#

These parameters change the shape of the response or refer to data stored at the provider, so they cannot be dropped silently:

The response carries the unsupported_parameter code and, in the param field, the parameter name or the path to the message part, e.g. messages[3].content[1]. Send the full conversation history with every request: the gateway keeps no provider-side state.