reasoning (the chain of thought) and content (the answer). The split means you can log, display, or discard the thinking without ever parsing it out of the reply:
reasoning arrives as deltas before content does, which makes long thinking phases feel alive instead of stalled.
Migrating? The trace used to arrive as
reasoning_content. That field is no longer emitted anywhere: read message.reasoning on a buffered reply and delta.reasoning on a streamed one. On the request side, the flat reasoning_effort is gone too, replaced by the reasoning object below.Reasoning effort
Every reasoning control rides one object,reasoning, with two members you can use today: enabled (a plain on/off) and effort. Effort is one of none, minimal, low, medium, high, xhigh, or max. Omit the object for the model’s default. In the OpenAI Python SDK pass it via extra_body:
How effort levels map
The seven levels are a normalized interface, we adopted the OpenRouter schema for them, so the same control works with every model whatever its worker natively calls its modes. Each model serves only some of those levels: thesupported_efforts list in the model’s reasoning block on GET /models is the served set, and the model table repeats it in its Reasoning column. Any level is accepted, and it resolves like this:
- A level the model serves is applied as is.
- Any other level rounds up to the next level the model serves, or to its highest level when you ask for more than it has.
noneis the disable control, not a rung: it switches thinking off on models not markedmandatoryand returns400on the ones that are, so you never get back reasoning tokens you didn’t expect.
low, medium, and xhigh, so minimal runs as low and high or max run as xhigh. GLM-5.3 serves low, high, and max, so medium runs as high and xhigh as max. DeepSeek V4 Pro 0813 serves high and max, so everything from minimal to medium runs as high: asking for low still succeeds even though its supported_efforts lists only max and high. The Responses API echoes the applied level in reasoning.effort, which is the quickest way to check.
Omitting the control gets the default_effort from the same block, and it varies: GLM-5.3 and GLM-5.3-Flash default to max, which thinks at length even on simple prompts. Coding agents and pipelines should set an effort explicitly rather than rely on the default.
reasoning.max_tokens (a thinking-only budget) returns 400 on every model today; each model’s reasoning block reports supports_max_tokens: false. To bound thinking, lower the effort or cap the top-level max_tokens, which covers the reasoning and the answer together.On Chat Completions, provider-native reasoning switches are rejected, not ignored: sending
reasoning_effort, thinking, chat_template_kwargs.enable_thinking, thinking_budget, or any of their siblings returns 400 on the key alone, whatever its value. Use reasoning. Clients that only know reasoning_effort need a custom body parameter instead; the integration guides show where each one sets it.Turning thinking off
Whether reasoning can be switched off entirely is per-model: it works only on models whosereasoning block on GET /models is not marked mandatory: true, and the reply then carries no reasoning and you stop paying for thinking tokens. Mandatory models always think and reject the disable with 400 rather than ignoring it; GLM-5.3 and GLM-5.3-Flash are the pair today, so budget for their thinking with effort and the top-level max_tokens instead. Either effort: "none" or enabled: false works. The example uses Qwen/Qwen3.8-27B, one of the non-mandatory models:
Messages and Responses
The same levels and the same mapping apply on the other API formats, through each format’s own field:- Anthropic Messages: effort goes in
output_config.effort, andthinking: {"type": "disabled"}turns thinking off. Thinking budgets (budget_tokens) are not supported. The trace comes back asthinkingcontent blocks. - OpenAI Responses: effort goes in
reasoning.effort, wherenoneturns thinking off. The reply echoes the applied level inreasoning.effort.
Multi-turn conversations
Treatreasoning as output for humans, not input for models: when you build the next turn’s history, append the assistant’s content only. The model re-thinks each turn, echoing old reasoning back just spends input tokens on text the model neither expects nor needs.
reasoning. An assistant message carrying reasoning_content is rejected with 400: presence is enough, so an explicit null fails too.
What thinking costs
- Reasoning tokens are output tokens, billed at the model’s output rate and counted inside
completion_tokens. Some workers itemize them ascompletion_tokens_details.reasoning_tokensin usage. - Thinking consumes the
max_tokensbudget before the answer does. If you seefinish_reason: "length"with lots of reasoning and a cut-off answer, raisemax_tokensor lower the reasoning effort. - For high-volume pipelines, benchmark GLM-5.3 Flash at
loweffort against DeepSeek V4 Pro 0813 at default, the answer-quality-per-dollar ordering depends on your task, and the playground makes the comparison quick.
Chain-of-thought is a working scratchpad, not a guaranteed explanation, audit decisions from the answer and your own evaluations, not from the narrative in
reasoning.