API Reference

Parameter Support by Model

These tables show which request parameters take effect for each model family. They are client-facing fields you send through any API format; the router translates each into whatever the model needs.

Legend: ✅ supported · ⚠️ conditional (see Per-Model Details) · — not supported (silently dropped).

A parameter not listed here is either dropped or forwarded unchanged to the model's backend, depending on the backend. A dropped parameter does not fail the request — the field simply has no effect — but a forwarded one that the model does not accept can return an error. Each parameter is described in the Parameters reference.

Core

Parameter	Type	GPT	Claude	Gemini	Other
`model`	string	✅	✅	✅	✅
`messages`	array	✅	✅	✅	✅
`temperature`	float	⚠️	⚠️	✅	✅
`top_p`	float	⚠️	⚠️	✅	✅
`max_tokens`	integer	✅	✅	✅	✅
`max_completion_tokens`	integer	✅	✅	✅	✅
`stop`	array	⚠️	✅	✅	⚠️
`stream`	boolean	✅	✅	✅	✅
`stream_options`	object	✅	✅	✅	✅

max_tokens is automatically converted to max_completion_tokens for GPT-5 models. GPT-5 and Codex models need max_tokens of at least 16.

stop is dropped for Grok models and for GPT-5. It must be an array of strings, even for a single sequence (for example ["END"]); a plain string is rejected with 400.

Claude: send only one of temperature and top_p. Claude 4.7 and later ignore temperature (other than 1), top_p and top_k, and with thinking on temperature is set to 1. See Per-Model Details.

Tool use

Parameter	Type	GPT	Claude	Gemini	Other
`tools`	array	✅	✅	✅	✅
`tool_choice`	string/object	✅	✅	✅	✅
`parallel_tool_calls`	boolean	✅	✅	✅	⚠️
`functions` (legacy)	array	✅	—	—	—
`function_call` (legacy)	string/object	✅	—	—	—

Reasoning & extended thinking

Parameter	Type	GPT	Claude	Gemini	Other
`thinking`	object	—	✅	✅	⚠️
`reasoning_effort`	string	✅	✅	✅	⚠️
`reasoning`	object	✅	✅	✅	⚠️

reasoning_effort, reasoning and thinking are normalised per model — the router maps whichever you send to the target model's native mode. reasoning is the nested form (effort, max_tokens, enabled, exclude); if both reasoning and reasoning_effort are sent, reasoning.effort wins. Other families support reasoning control only on models with a switchable reasoning mode. See Per-Model Details and Reasoning & Thinking.

Output control

Parameter	Type	GPT	Claude	Gemini	Other
`response_format`	object	✅	✅	✅	✅
`n`	integer	✅	—	✅	—
`seed`	integer	✅	—	✅	—
`user`	string	✅	✅	✅	✅
`web_search_options`	object	⚠️	✅	✅	—

web_search_options works on GPT models with web search support and has no effect on others, such as gpt-4.1.

Sampling & penalties

Parameter	Type	GPT	Claude	Gemini	Other
`frequency_penalty`	float	⚠️	—	⚠️	✅
`presence_penalty`	float	⚠️	—	⚠️	✅
`logit_bias`	object	⚠️	—	—	—

GPT-5 models always drop both penalties and logit_bias; earlier GPT models accept them. Penalties are not available on Gemini preview and Gemini 3 models.

Logprobs

Parameter	Type	GPT	Claude	Gemini	Other
`logprobs`	boolean	⚠️	—	✅	⚠️
`top_logprobs`	integer	⚠️	—	✅	⚠️

Logprobs apply to Gemini models and to GPT models. GPT-5.1, 5.2, 5.4 and 5.5 return them only when reasoning_effort is "none"; other GPT-5 models (including Codex variants) drop them. For other model families, support varies by model. Claude models ignore them.

Extended parameters

These are not top-level fields. Pass them inside extra_body; each is extracted by an allowlist before being merged into the outbound request. (service_tier is the exception — it is a standard top-level field.)

Parameter	Via	GPT	Claude	Gemini	Other
`modalities`	`extra_body`	✅	—	—	—
`audio`	`extra_body`	✅	—	—	—
`prediction`	`extra_body`	✅	—	—	—
`service_tier`	top-level	✅	—	—	—
`top_k`	`extra_body`	—	⚠️	✅	⚠️

top_k is not part of the standard request body — pass it via extra_body; it is forwarded only to models that support it natively. Claude 4.7 and later ignore it.