UltraThink
Extended thinking in Claude Code, the ultrathink keyword, effort level, and how to tune thinking tokens.
UltraThink is not a separate product: it is how you ask Claude Code to reason harder on a turn using extended thinking, /effort, toggles, and the ultrathink keyword.
Official background: Common workflows (extended thinking) and Extended thinking on the Claude platform.
How extended thinking works in Claude Code
Claude Code enables extended thinking by default. On supported models, effort level controls adaptive reasoning so routine prompts stay fast and harder steps get more depth.
The ultrathink keyword is a prompt-level nudge: include the word ultrathink anywhere in your message and Claude gets an in-context instruction to reason more on that turn. It does not change your saved effort level. (Casual phrases like “think hard” are not treated as special token switches.)
Ways to tune thinking
Summarized from Common workflows:
| Approach | What to do |
|---|---|
| Effort level | Run /effort, adjust in /model, or set CLAUDE_CODE_EFFORT_LEVEL. |
ultrathink keyword | Put ultrathink in the prompt for extra reasoning on that turn only. |
| Session toggle | Option+T (macOS) or Alt+T (Windows/Linux) toggles thinking for the session. |
| Global default | /config (stored as alwaysThinkingEnabled in ~/.claude/settings.json). |
| Token budget cap | MAX_THINKING_TOKENS (see Environment variables). Positive values apply only to fixed-budget models; adaptive models ignore them. 0 disables thinking on the Anthropic API except for Fable, where thinking cannot be disabled. |
Press Ctrl+O for verbose mode if you want to see reasoning-style output in the UI (details depend on your settings and model).
Fable 5-family models, Sonnet 5, and Opus 4.7 or later always use adaptive reasoning. Opus 4.6 and Sonnet 4.6 can return to a fixed budget with CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1.
Example prompt
ultrathink: we can implement rate limiting as a token bucket or a fixed window on this endpoint. Compare ops tradeoffs for a single-region deploy, pick one, and list the exact files or modules you would touch first, no full implementation yet.Skills
In skills documentation, including the word ultrathink in a skill’s content can enable extended thinking when that skill runs, so heavier tasks get more internal reasoning without you typing the keyword every time. See Skills.