Parameters

Send parameters in the JSON body of POST /v1/chat/completions. Parameters are forwarded to the upstream provider as-is, so support is provider- and model-dependent — a model that doesn't recognise one ignores it.

model#

The model to run. Use a provider/model slug from the model catalog, for example anthropic/claude-opus-4.8 or openai/gpt-5.4. The full list with DA pricing is at GET /api/models.

messages#

The conversation history as an ordered array of message objects. Each object has a role (system, user, assistant, or tool) and a content field. Content may be a plain string or an array of content parts (text and/or image URLs) for vision-capable models.

[
  { "role": "system", "content": "You are a helpful assistant." },
  { "role": "user", "content": "What is the cost of 1 million tokens?" }
]

See Messages & roles for the full content-parts schema.

models#

Type: array of strings · Optional · max 4 entries

Fallback chain. If the primary model errors or is rate-limited upstream, the request is retried with the next slug in this list, in order. Billing always reflects the model that actually served the response.

{
  "model": "anthropic/claude-opus-4.8",
  "models": ["openai/gpt-5.4", "google/gemini-2.5-pro"],
  "messages": [{ "role": "user", "content": "Hello" }]
}

Provider-level failover within a single model happens automatically and does not need this parameter.

stream#

Set to true to receive tokens as server-sent events instead of waiting for the full completion. See Streaming.

max_tokens#

Maximum number of tokens the model may generate. Setting this caps spending on a single request. If the model hits this limit, finish_reason will be "length". The OpenAI-style alias max_completion_tokens is also accepted and forwarded.

temperature#

Sampling temperature. Lower values (e.g. 0.2) make output more deterministic; higher values (e.g. 1.4) increase randomness. For tasks requiring reproducible outputs, use 0. Cannot be set simultaneously with top_p on some models.

top_p#

Nucleus sampling threshold. The model samples from the smallest set of tokens whose cumulative probability exceeds top_p. 1.0 means no filtering. OpenAI recommends altering either temperature or top_p, but not both.

The following parameters are forwarded to the upstream provider as-is. Support is provider- and model-dependent — a model that doesn't recognise one ignores it.

top_k#

Limits sampling to the top-K most probable tokens at each step. Not exposed by all providers.

frequency_penalty#

Reduces the likelihood of tokens that have already appeared frequently in the output. Positive values discourage repetition.

presence_penalty#

Reduces the likelihood of tokens that have appeared at all in the output so far. Encourages introducing new topics.

repetition_penalty#

Alternative repetition control used by some providers. Values above 1.0 penalize repetition; values below 1.0 encourage it.

min_p#

Minimum probability threshold relative to the top token. Tokens with probability below min_p × p(top token) are discarded.

top_a#

Alternative nucleus filter based on the absolute top-token probability. Tokens with probability below top_a × p(top token)² are discarded.

seed#

If set, the model attempts deterministic sampling. Same seed + same prompt should produce the same output (best-effort; not guaranteed across versions).

logit_bias#

Map of token IDs to bias values (-100 to 100). Use to ban specific tokens (-100) or boost them (100).

logprobs#

Return log probabilities for output tokens when true.

top_logprobs#

Number of most-likely tokens to return log probabilities for at each position. Requires logprobs: true.

response_format#

Controls output format. Set { "type": "json_object" } to request guaranteed JSON output (structured outputs). See Features for details.

stop#

One or more sequences at which the model will stop generating. The stop string itself is not included in the output.

tools#

Array of tool definitions the model may call. Each tool has a type ("function"), a name, a description, and a JSON Schema parameters object.

tool_choice#

Controls whether the model calls a tool. "auto" (default), "none", "required", or a specific { "type": "function", "function": { "name": "..." } }.

parallel_tool_calls#

Set to false to force the model to call at most one tool per turn.

reasoning#

The unified control for extended chain-of-thought reasoning on models that support it. Accepts:

• effort — "high", "medium", "low" (some models also accept "max", "xhigh", "minimal", "none").
• max_tokens — an explicit thinking-token budget, for models that support direct allocation. Use effort or max_tokens, not both.
• exclude — true makes the model reason internally without returning the reasoning text.
• enabled — true turns reasoning on with the model's default settings.

Reasoning tokens are billed at the model's output token rate. Models that return their reasoning expose it on the assistant message as a reasoning string (or a structured reasoning_details array).

{
  "model": "anthropic/claude-opus-4.8",
  "messages": [{ "role": "user", "content": "Solve this integral step by step." }],
  "reasoning": { "effort": "medium" }
}

Models that support reasoning carry the reasoning capability flag in the catalog — filter with GET /api/models?capability=reasoning.

reasoning_effort#

The OpenAI-style scalar alias for reasoning.effort (e.g. reasoning_effort: "high"). Use it when a client SDK only exposes the flat field; the structured reasoning object is the recommended form.