Skip to content

Instantly share code, notes, and snippets.

@intellectronica
Created November 25, 2025 13:57
Show Gist options
  • Select an option

  • Save intellectronica/58571dda3581eec3e17a77741e8c858a to your computer and use it in GitHub Desktop.

Select an option

Save intellectronica/58571dda3581eec3e17a77741e8c858a to your computer and use it in GitHub Desktop.
Claude Extended Thinking: The Ultimate Guide

Claude Extended Thinking: The Ultimate Guide

When to use extended thinking across Claude models in the Claude App and Claude Code

Generated by Claude Opus 4.5 (with extended thinking)


What Is Extended Thinking?

Extended thinking gives Claude a "scratchpad" to reason through problems before responding. It's the same model with more time to deliberate — not a separate model. Performance on complex tasks improves logarithmically with thinking tokens allocated.

Key insight: Extended thinking isn't universally better. Research shows it can hurt performance by up to 36% on certain task types — similar to how humans perform worse when overthinking intuitive tasks.


Quick Reference Table

App Model Default When to Enable When to Keep Off
Claude App Sonnet 4.5 Off Complex reasoning, debugging, analysis Simple Q&A, creative writing, lookups
Claude App Opus 4.5 On (Max) Leave on for most tasks Quick chat, pattern recognition
Claude App Haiku 4.5 Off Hard problems where speed matters less Real-time tasks, rapid iteration
Claude Code Sonnet 4.5 Off Architecture, complex bugs, planning Routine edits, boilerplate
Claude Code Opus 4.5 Off System design, multi-codebase work Most tasks — Opus is already strong
Claude Code Haiku 4.5 Off When stuck in loops, multi-step logic Prototyping, scaffolding

Claude App (claude.ai / Desktop / Mobile)

How to Enable

  1. Click "Search and tools" button (lower left)
  2. Toggle "Extended thinking" on

⚠️ Toggling mid-conversation starts a new chat.

Plan Access

Plan Models Available
Free Sonnet 4.5, Haiku 4.5
Pro Sonnet 4.5, Haiku 4.5
Max All models (Opus 4.5 is default)

Sonnet 4.5

Use thinking for:

  • Multi-step math or logic problems
  • Debugging complex code
  • Architectural planning
  • Legal/financial analysis
  • When accuracy matters more than speed

Keep thinking off for:

  • Simple factual questions
  • Quick translations
  • Summarisation
  • Creative writing (style/voice)
  • Pattern recognition tasks

Opus 4.5

Opus is a hybrid model — reasons deeply even without extended thinking. For Max users, defaults to thinking on.

Use thinking for:

  • Leave it on for most tasks
  • Complex multi-step analysis
  • Deep research and synthesis

Keep thinking off for:

  • Quick chat, simple Q&A
  • When latency matters
  • Pattern matching / classification

Haiku 4.5

First Haiku with extended thinking. Use when you need speed + occasional depth.

Use thinking for:

  • Complex coding problems
  • Multi-step reasoning
  • When stuck on a hard problem

Keep thinking off for:

  • Quick lookups, simple chat
  • Real-time interactions
  • High-throughput tasks

Claude Code

How to Enable

Method 1: Tab key

  • Press Tab to toggle thinking on/off
  • Sticky across sessions

Method 2: Keywords in prompts

Keyword Tokens Use Case
think ~4,000 Routine debugging, basic refactoring
think hard / megathink ~10,000 Architectural decisions, complex problems
think harder / ultrathink ~32,000 Hardest tasks, deep sustained reasoning

Plan Access

Plan Models Available
Pro Sonnet 4.5, Haiku 4.5 (no Opus)
Max All models including Opus 4.5

Recommended Workflow

1. EXPLORE    → Let Claude read files (don't write code yet)
2. PLAN       → "ultrathink. Analyse this and propose a plan. Don't code yet."
3. CODE       → Implement based on confirmed plan
4. COMMIT     → Ask Claude to commit and create PR

Steps 1–2 are crucial. Without them, Claude jumps straight to coding.

Sonnet 4.5 (Default)

Trigger thinking for:

  • ultrathink before architectural decisions
  • think hard for complex debugging
  • When Claude gets stuck in a loop
  • Multi-file refactoring
  • Before writing code — ask for a plan first

No keyword needed for:

  • Simple file edits
  • Routine refactoring
  • Generating boilerplate
  • Quick questions about code

Opus 4.5 (Max Only)

Trigger thinking for:

  • Complex system design
  • Multi-codebase refactoring
  • Debugging subtle concurrency issues

No keyword needed for:

  • Most tasks — Opus is already strong
  • Simple changes
  • When speed matters

Note: Opus 4.5 uses fewer tokens than Sonnet to solve the same problems.

Haiku 4.5

Trigger thinking for:

  • When Claude gets stuck
  • Complex logic problems
  • Multi-step analysis

No keyword needed for:

  • Rapid prototyping
  • UI scaffolding
  • Simple file operations

When Extended Thinking Hurts Performance

Research shows thinking can degrade performance on these task types:

Task Type Why Thinking Hurts
Pattern recognition Intuition beats deliberation
Visual classification Statistical learning is implicit
Classification with exceptions Overthinking misses edge cases
Implicit learning tasks Analysis interferes with pattern matching

Rule of thumb: If a human would do worse by "thinking too hard," so will Claude.


Cheat Sheet

Claude App — Quick Decision

Simple question      → OFF
Complex reasoning    → ON
Creative writing     → OFF
Code debugging       → ON
Quick lookup         → OFF
Analysis/research    → ON
Pattern matching     → OFF

Claude Code — Quick Decision

Quick fix            → No keyword
Routine refactor     → "think"
Architecture         → "ultrathink" + plan first
Stuck in loop        → "ultrathink"
Simple edit          → No keyword
New feature          → "think hard" + plan first
Prototyping          → No keyword

Model Comparison

Metric Haiku 4.5 Sonnet 4.5 Opus 4.5
SWE-bench 73.3% 77.2% 80.9%
Speed Fastest (2x Sonnet) Fast Slower
Best for High-volume, real-time Daily development Complex/critical tasks
Extended thinking First Haiku with it Recommended for coding Usually unnecessary

Key Takeaways

  1. Start without thinking — enable when you hit problems needing deeper analysis
  2. Plan before coding — especially with ultrathink in Claude Code
  3. Opus rarely needs thinking — it's already a strong reasoner
  4. Haiku + thinking — unlocks near-Sonnet performance when needed
  5. Avoid thinking for pattern tasks — it can hurt performance significantly
  6. Keywords only work in Claude Code — not in the Claude App or API

Last updated: November 2025

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment