Reasoning effort

Reasoning effort controls how much thinking the model applies to your prompts. Lower effort levels produce faster, shorter responses and use fewer credits. Higher levels spend more tokens on deeper analysis, multi-step reasoning, and thorough code generation.

Available levels:

Not all models support every level. The picker only shows levels available for your current model - see supported models below.

Each model defines the values it accepts for the reasoning-related fields in its configuration schema. The output_config.effort column lists the effort levels available for that model; thinking.type and thinking.display control reasoning behavior and visibility; max_tokens sets the output length limit.

Claude models use output_config.effort with thinking.type and thinking.display controls:

GPT-5.6 models use reasoning.effort:

GPT-5.6 models default to effort: "high".

Your effort choice persists automatically. Set a level with /effort or --effort and Kiro remembers it for future sessions - there's no extra step to make it the default. Preferences are stored in ~/.kiro/settings/cli.json. See In-session settings for more on how preferences persist.

To set default model parameters per model - so you don't have to run /effort at the start of every session - add chat.modelDefaults to your settings file:

Claude models (use output_config.effort):

{
  "chat.modelDefaults": {
    "claude-sonnet-4.6": {
      "output_config": {
        "effort": "high"
      }
    },
    "claude-opus-4.7": {
      "output_config": {
        "effort": "max"
      }
    }
  }
}

GPT-5.6 models (use reasoning.effort):

{
  "chat.modelDefaults": {
    "gpt-5.6-sol": {
      "reasoning": {
        "effort": "max"
      }
    },
    "gpt-5.6-terra": {
      "reasoning": {
        "effort": "low"
      }
    }
  }
}

Control whether the model uses extended thinking and how it displays reasoning:

{
  "chat.modelDefaults": {
    "claude-opus-4.8": {
      "thinking": {
        "type": "adaptive",
        "display": "summarized"
      }
    },
    "claude-sonnet-4.6": {
      "thinking": {
        "type": "disabled"
      }
    }
  }
}

• thinking.type - adaptive enables extended thinking when the model determines it's needed; disabled turns it off entirely.
• thinking.display - summarized shows a condensed version of reasoning; omitted hides it from output. Only applies when type is adaptive.

Set the maximum number of tokens the model can generate per response:

{
  "chat.modelDefaults": {
    "claude-opus-4.8": {
      "max_tokens": 128000
    },
    "claude-sonnet-4.6": {
      "max_tokens": 32000
    }
  }
}

max_tokens limits per model:

You can combine output_config, thinking, and max_tokens in a single model entry:

Claude example:

{
  "chat.modelDefaults": {
    "claude-opus-4.8": {
      "output_config": {
        "effort": "max"
      },
      "thinking": {
        "type": "adaptive",
        "display": "summarized"
      },
      "max_tokens": 128000
    }
  }
}

GPT-5.6 example:

{
  "chat.modelDefaults": {
    "gpt-5.6-sol": {
      "reasoning": {
        "effort": "max"
      },
      "max_tokens": 128000
    }
  }
}

To open your settings file in your editor:

kiro-cli settings open

Or place a .kiro/settings/cli.json in your project root to set workspace-level defaults that apply to everyone working in that repository.

When determining the effort level for a session, Kiro applies this priority order:

1. Session override - value set via the effort picker, /effort, or --effort during the current session
2. Workspace defaults - chat.modelDefaults in .kiro/settings/cli.json
3. User defaults - chat.modelDefaults in ~/.kiro/settings/cli.json
4. Built-in defaults - the model's standard effort level

• Bump up when the agent is giving shallow answers, missing edge cases, or producing incomplete implementations
• Bump down when you need quick answers and don't want to wait for extended reasoning
• Use max for security reviews, complex debugging sessions, or when you need the agent to consider many interacting constraints

• Models - available models and their capabilities
• Slash commands reference - quick command reference
• Settings - all configurable settings