Reasoning

Configure reasoning effort and reasoning output only on models that expose those controls.

Overview

Reasoning (or "thinking") models spend extra internal tokens working through a problem before they answer, which improves accuracy on hard maths, coding, and multi-step tasks. You control how much effort the model spends; the thinking itself may be summarized or hidden depending on the model. Reasoning is opt-in and only applies to models that support it.

Common controls

ParameterTypeRequiredDescriptionValues
reasoning_effortChatoptionalOpenAI-style reasoning effort field: `low`, `medium`, or `high`.
lowmediumhigh
reasoning.effortResponsesoptionalSame control on the Responses API, nested under `reasoning`.
thinkingClaudeoptionalClaude extended-thinking config, e.g. `{"type":"enabled","budget_tokens":2048}`.
thinkingConfigGeminioptionalGemini native thinking configuration when supported.

Example

from openai import OpenAI

client = OpenAI(base_url="https://onerouter.com/v1", api_key="YOUR_API_KEY")

completion = client.chat.completions.create(
    model="openai/gpt-5.2",
    reasoning_effort="high",
    messages=[{"role": "user", "content": "If 3 painters paint 3 fences in 3 hours, how many painters paint 9 fences in 9 hours?"}],
)

print(completion.choices[0].message.content)

Reasoning tokens in usage

When a model reasons, the work shows up in usage as reasoning/thinking tokens, billed separately from visible output. Watch these to keep cost predictable.

JSON
{
  "usage": {
    "prompt_tokens": 32,
    "completion_tokens": 180,
    "completion_tokens_details": { "reasoning_tokens": 140 },
    "total_tokens": 212
  }
}

Billing impact

Reasoning models may bill reasoning tokens or internal computation separately from visible output tokens. Review the model pricing before increasing reasoning effort by default.