Reasoning

Thinking models, the reasoning object, and how thought is billed.

Models tagged REASONING in the catalog can think before answering. Request it with the reasoning object — one surface across every provider, whatever their native API calls it:

{
  "model": "openai/gpt-oss-120b",
  "stream": true,
  "reasoning": { "effort": "low" },
  "messages": [{ "role": "user", "content": "Which is larger, 9.11 or 9.9?" }]
}

Field	Meaning
`effort`	`minimal`, `low`, `medium`, or `high` — mapped to each provider's dial
`max_tokens`	explicit thinking budget, for budget-style models
`exclude`	`true` = think internally but omit the text from the response
`enabled`	`false` = disable thinking on models that default to it

Reading the thought stream

Reasoning text arrives on its own delta field, separate from content:

data: {"choices":[{"delta":{"reasoning":"Compare decimals: 9.11 vs 9.9 …"},…}]}
data: {"choices":[{"delta":{"content":"9.9"},…}]}

Some models attach reasoning_details entries alongside — summaries, provider signatures, or encrypted blocks that let multi-turn tool use resume correctly. Pass them back untouched if you continue the conversation.

Whether the thought TEXT is visible depends on the model and route; thinking TOKENS are always accounted either way.

Thinking variants

Some models also exist as a :thinking catalog variant (for example anthropic/claude-opus-5:thinking) — same underlying model with thinking pre-enabled and its own pricing row. Using the base id with a reasoning object and using the variant are equivalent; the variant exists so a plain model switch is enough for clients that can't send extra fields.

Billing

Thinking tokens are reported in usage.completion_tokens_details.reasoning_tokens, inside completion_tokens. They're billed at the model's reasoning rate when one is listed, otherwise at the output rate — and the in-band usage.cost already includes them, so there is never a surprise.