Models & Thinking Modes

The Three Models

Sonnet is the natural starting point, covering 80% of use cases. Haiku cuts the cost by 4x on repetitive tasks (test generation, mass renaming). Opus only comes into play when deep reasoning genuinely changes the quality of the result.

Switching Models

# Via slash command
/model

# Via flag at startup
claude --model haiku "Fix this typo in README.md"
claude --model opus "Design the auth system"

Quick Decision Table

Thinking Modes (Opus 4.6)

Opus 4.6 uses Adaptive Thinking: the model dynamically allocates its compute budget based on request complexity. The effort parameter controls this depth.

Effort levels:

• low: zero preamble, combined operations, mechanical tasks
• medium: defined pattern, bounded scope, single concern
• high: design decisions, edge cases, multiple concerns
• max: cross-system reasoning, irreversible decisions (Opus 4.6 only)

Controlling Thinking

Alt+T          Toggle on/off (current session)
/config        Persist across sessions
/model + ←/→  Effort slider low|medium|high

The keywords ultrathink and think hard have had no functional effect since v2.0.67. Adaptive thinking is active by default on Opus 4.6.

Golden Rule

Start with Sonnet. Move up to Opus only if the task involves irreversible decisions, multi-system reasoning, or a critical audit. Drop to Haiku for anything mechanical and predictable. Most sessions never need Opus.

Multi-Agent Pattern

# planner.md — exploration, read-only
model: opus
tools: Read, Grep, Glob

# implementer.md — mechanical execution
model: haiku
tools: Write, Edit, Bash

Assign models based on role, not perceived importance.