Overview
Reasoning models can process information in two stages:- Thinking stage: Internal reasoning process, returned in the
reasoning_contentfield of the response message - Response stage: The final answer, returned in the
contentfield
chat_template_kwargs parameter in your API requests. Reasoning output is billed as completion tokens. The number of tokens spent thinking is reported in usage.completion_tokens_details.reasoning_tokens, the same location OpenAI uses, and is a subset of completion_tokens. The top-level usage.reasoning_tokens field is kept as a deprecated alias.
GLM-5.3 Flash
GLM-5.3 Flash (z-ai/glm-5.3-flash) reasons by default. The reasoning is returned in reasoning_content.
Cap the reasoning with thinking_token_budget
thinking_token_budget sets the maximum number of reasoning tokens. When you omit it, the budget is 8192 tokens. Send null to remove the cap. This is the most precise control for this model:
reasoning_effort is close to on/off
GLM-5.3 Flash accepts reasoning_effort, but the levels are not a gradual dial:
To shorten the reasoning without removing it, use
thinking_token_budget instead.
GLM-5.1
GLM-5.1 supports reasoning through theenable_thinking parameter in chat_template_kwargs. Reasoning is enabled by default for GLM-5.1.
Default Behavior (Reasoning Enabled)
GLM-5.1 uses reasoning by default. You can explicitly enable it by setting"enable_thinking": true, or simply omit the parameter:
Disable Reasoning
To disable reasoning for GLM-5.1, you must explicitly set"enable_thinking": false in the chat_template_kwargs:
Qwen3.5-122B and Qwen3.6-35B
Qwen3.5-122B (Qwen/Qwen3.5-122B-A10B) and Qwen3.6-35B (Qwen/Qwen3.6-35B-A3B-FP8) have reasoning enabled by default. The model will include reasoning_content in responses automatically. To disable reasoning, set "enable_thinking": false in the chat_template_kwargs:
GPT-OSS-120B
GPT-OSS-120B (openai/gpt-oss-120b) always reasons before answering — its reasoning cannot be disabled. Instead, you control how much it thinks with the standard OpenAI reasoning_effort parameter (low, medium, or high; defaults to medium):
Kimi K2.5 and Kimi K2.6
Kimi K2.5 (moonshotai/kimi-k2.5) and Kimi K2.6 (moonshotai/kimi-k2.6) have reasoning enabled by default and include reasoning_content in responses automatically. To disable reasoning, set "thinking": false in the chat_template_kwargs (note the parameter name — thinking, not enable_thinking):
"thinking": false the response contains no reasoning trace and usage.completion_tokens_details.reasoning_tokens drops to 0.
Model-Specific Parameters
Different models use different parameter names for controlling reasoning:
Third-party models served through the gateway (OpenAI
o-series and GPT-5.x, Anthropic Claude, Gemini Pro, …) use their provider’s standard reasoning controls, such as reasoning_effort. The Kimi models are the exception: they do not respond to reasoning_effort — use chat_template_kwargs.thinking as described above.
Reading Reasoning Output
When reasoning is active, responses contain the reasoning trace alongside the answer:reasoning_content— the model’s internal thinking (streamed asdelta.reasoning_contentchunks whenstream: true)usage.completion_tokens_details.reasoning_tokens— how many of the completion tokens were spent thinking. It is present in non-streaming responses and in the final usage chunk of a stream when you setstream_options: {"include_usage": true}. The top-levelusage.reasoning_tokenscarries the same number and is deprecated.
When to Use Reasoning
Reasoning is particularly useful for:- Complex mathematical problems: Multi-step calculations and problem-solving
- Logical reasoning: Tasks requiring step-by-step analysis
- Code generation: Complex programming problems that need careful planning
- Scientific questions: Problems requiring structured thinking
- Simple queries: Straightforward questions with direct answers
- Fast responses: When latency is critical and the problem is simple
- Cost optimization: Reasoning can increase token usage and costs
Best Practices
- Test with and without reasoning: Compare results to determine if reasoning improves output quality for your use case
- Monitor token usage: Reasoning can increase the number of tokens used, affecting costs — watch
usage.reasoning_tokens - Use appropriate models: Not all models support reasoning - check model capabilities before enabling
- Stream responses: When using reasoning, streaming can provide better user experience for longer responses