Skip to main content
Configure reasoning to improve model performance on complex tasks.
Reasoning-capable language models can spend additional computation to “think” through problems step-by-step before generating a final response. This can improve performance on complex tasks. You can control this behavior using the reasoning_effort parameter. Available values and default behavior depend on the selected model. Check the Reasoning column under Available Models to see which models support this feature.

Configure reasoning effort

You can configure reasoning_effort in your API request, which controls how much reasoning the model performs. If not manually set, it will run with its default value, which is model-dependent.

Choosing reasoning effort level

Consider your task’s complexity, your application’s latency, and your token budget. Higher effort can improve performance on complex tasks, but can also increase latency and output token consumption.
  • Default effort: Use the model’s default unless you have established that another setting performs better for your use case.
  • Lower effort: Best-suited for simple tasks and latency-sensitive requests.
  • Higher effort: Best-suited for complex debugging, mathematical problems, multi-step analysis, and planning.
  • None: Disables thinking, making the model behave like a standard chat model.
Models use internal reasoning_tokens before producing a response. These are billed and counted as output tokens.

Model-specific behavior

Available reasoning effort values and default behavior varies between models. See the model-specific documentation for more details.
ModelAvailable ValuesDefault Value
DeepSeek V4.1 Flash 1017none, low, high, maxhigh
GLM 5.3 Flashlow, high, maxhigh