Create a chat completion

Chat completions (LLM). Set stream: true for token-by-token SSE streaming.

Body

modelstringYes

Model ID (e.g. everyais/claude-opus-5)

messagesarrayYes

Array of 1–1000 messages. role is system / user / assistant / tool / developer (normalized to system). content is a string, null, or an array of multimodal parts (up to 1000)

streambooleanNo

SSE streaming (default false)

stream_optionsobjectNo

With {"include_usage": true}, the final chunk includes usage

max_tokensintegerNo

Maximum output tokens, 1–200000. max_completion_tokens is normalized automatically

temperaturenumberNo

0–2

top_pnumberNo

0–1

stopstring | string[]No

Up to 4

nintegerNo

Only 1 is supported (default 1). 2 or more returns 400

presence_penalty / frequency_penaltynumberNo

-2–2

seedintegerNo

Reproducibility hint

toolsarrayNo

Function tool definitions (up to 512) and the server-side web search {"type":"web_search"} (up to 1 — see the web search guide)

tool_choicestring | objectNo

auto / none / required, or {"type":"function","function":{"name":"..."}}. none also disables web search

response_formatobjectNo

{"type":"text"} · {"type":"json_object"} · {"type":"json_schema","json_schema":{...}}

reasoning_effortstringNo

none / low / medium / high

parallel_tool_callsbooleanNo

Allow parallel tool calls

logprobs / top_logprobsboolean / integerNo

Forwarded on Chat Completions routes that support them. top_logprobs is 0–20. Explicit values return 400 on OpenAI/Mantle upstream Responses routes

userstringNo

End-user identifier

everyaisobjectNo

Gateway options — {"cache":"on"|"off","cache_ttl":"5m"|"1h"} and provider

providerobjectNo

Routing hints: sort (price / latency / throughput) and allow_fallbacks. Selecting providers by name is not supported.

modelsstring[]No

Catalog fallback, max 5. Tried before reservation when the first model is not callable.

extra_bodyobjectNo

Provider-specific extensions — only the anthropic / google / openai / everyais keys are allowed (any other key returns 400). extra_body.provider is rejected.

You may append :nitro (prefer lower latency) or :floor (prefer lower endpoint cost) to a model slug. Billing does not change.

Undefined OpenAI parameters (logit_bias, store, etc.) are silently ignored.

Model-specific options

The table above describes the shared input schema, not universal model support. Check capabilities and limits in GET /v1/models. false means unsupported; an omitted field means support is unverified. Explicit unsupported options, values, and combinations return 400.

• With thinking enabled on Claude 4.6 and earlier, explicit temperature must be 1, top_p must be 0.95–1, and top_k returns 400. Opus/Sonnet 4.5 and 4.6, plus Haiku 4.5, also reject specifying both temperature and top_p. Manual thinking cannot be combined with tool_choice: "required" or a named function. Fable/Mythos 5.1 does not support forced tool selection at all.
• Chat max_tokens can be capped at the model's output limit. A manual budget supplied directly through extra_body.anthropic.thinking is not added to max_tokens; it must be smaller than max_tokens.
• An OpenAI-compatible request format does not imply identical options across provider routes. OpenAI/Mantle upstream Responses routes (including GPT-6 and Responses-only models) return 400 for explicit stop, seed, presence_penalty, frequency_penalty, logprobs, or top_logprobs. Even logprobs: false differs from omission. These options are not silently ignored; user is forwarded unchanged as the end-user identifier. This provider-route restriction also applies when your public endpoint is /v1/chat/completions.
• capabilities.sampling describes sampling support; limits.reasoning_efforts lists the model's reasoning levels. Omission uses the model default and differs from none.
• json_mode means JSON output requests; structured_outputs means native strict schema support. Claude json_object and JSON schemas with strict omitted or false use a best-effort prompt and do not guarantee schema compliance. strict: true requires native support; otherwise it returns 400. Claude function.strict also requires native support.
• Strict function tools have separate support requirements from JSON schemas. Gemini 3 function.strict: true uses the provider's VALIDATED mode for automatic tool selection. Gemini 2.5 supports JSON schemas but rejects strict function tools (strict: true) with 400.
• Claude native strict support depends on the route. Bedrock supports Opus/Sonnet 4.5 and 4.6, plus Haiku 4.5; this does not automatically extend to Opus 5. Rely on support only when the public capability is true.
• Claude honors parallel_tool_calls: false by disabling parallel tool use. Gemini rejects false with 400; true or omission allows the provider's parallel calls.
• Gemini 2.5 rejects combining JSON output (json_object/json_schema) with tools, and rejects mixing web search with function tools, with 400. Use those features separately. Gemini 3 supports these combinations. capabilities.json_mode_with_tools and web_search_with_tools are false for unsupported combinations; omission means unverified support.
• In a function-tool loop, replay the original assistant tool_calls unchanged to the same model, followed by the tool result for its tool_call_id. Gemini tool_calls[].extra_content.google.thought_signature is opaque provider state; do not modify it, strip it, or reuse it with another model.
• web_search is available only on Chat models with capabilities.web_search: true. Gemini tool_choice: "none" also disables search. Requiring a tool or naming a function when only a search tool is present returns 400.

Streaming

With stream: true, data: {...} SSE chunks follow one another and end with data: [DONE]. Enabling stream_options.include_usage appends a final usage chunk.

See the streaming guide for a complete example and completion/error handling.