Models & Thinking Modes — Claude Code Recap Card T18

The Models

Model	Input price/1M	Ideal for
Haiku 4.5	$0.80	Mechanical tasks, CI/CD
Sonnet 4.6	$3.00	Day-to-day development (default)
Opus 4.6	$5.00	Architecture, security, audits
Opus 4.7	—	Most capable model; xhigh effort available

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. Opus 4.7 (claude-opus-4-7) is the most powerful option and is the only model with xhigh effort support.

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

Task	Model	Effort
Renaming, formatting	Haiku	low
Test generation	Haiku	low
Standard feature	Sonnet	medium
Module refactoring	Sonnet	high
System architecture	Opus	high
Critical security audit	Opus	max

Thinking Modes (Opus 4.6 / 4.7)

Opus 4.6 and 4.7 use Adaptive Thinking: the model dynamically allocates its compute budget based on request complexity. The effort parameter controls this depth.

Effort levels (v2.1.111+):

xlow: absolute minimum, speed priority
low: zero preamble, combined operations, mechanical tasks
default: balanced (applies when no effort is set)
high: design decisions, edge cases, multiple concerns — default on Pro and Max plans since v2.1.117
max: cross-system reasoning, irreversible decisions (Opus 4.6)
xhigh: maximum depth, Opus 4.7 only

Controlling Thinking

Alt+T          Toggle on/off (current session)
/config        Persist across sessions
/effort        Set effort level interactively
/model + ←/→  Effort slider

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 and 4.7.

CI/CD: use the CLAUDE_EFFORT env var to pin effort level in automated pipelines (v2.1.120):

CLAUDE_EFFORT=high claude -p "Review this PR" --dangerously-skip-permissions

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.