Claude Code ultrathink Guide 2026: Trigger AI Deep Thinking Mode

Applies to: Claude Code v2.1.72+ (March 2026)

How to Use ultrathink (30-Second Version)

Add ultrathink at the start of your prompt:

ultrathink analyze the architectural bottleneck here and propose three refactoring approaches

ultrathink I've been debugging this for two hours with no luck — do a deep analysis of all possible causes

This triggers Claude Code's high effort mode, giving the AI maximum reasoning resources for that turn.

Key point: ultrathink only affects the current turn. After it completes, Claude automatically returns to the default medium effort.

ultrathink Version History (2025 → 2026)

Many older articles still say "ultrathink = 31,999 tokens" — that was the v1 behavior and no longer applies. The current ultrathink triggers an effort level, not a fixed token count.

The 2026 Effort System

ultrathink vs Permanent Effort Setting

• ultrathink → high effort for that one turn only, then reverts to medium — best for occasional use
• /model command → sets effort level for the entire session — better when you need deep thinking throughout a working session

Official Verification

This is not a community-discovered workaround — it's an officially documented Claude Code feature.

From the Claude Code Best Practices official docs:

"We recommend using the word 'think' to trigger extended thinking mode, which gives Claude additional computation time to evaluate alternatives more thoroughly. These specific phrases are mapped directly to increasing levels of thinking budget in the system: 'think' < 'think hard' < 'think harder' < 'ultrathink.'"

This describes the v1 mechanism. In v2, Anthropic switched to the effort system — and officially re-introduced the ultrathink keyword in v2.1.68 as a shortcut to trigger high effort. The underlying implementation changed, but the intent is identical.

When to Use ultrathink

Good Use Cases

• Complex architecture decisions — weighing multiple design approaches with real trade-offs
• Hard-to-trace bugs — when you've tried multiple things and still can't find the root cause
• Deep code analysis — understanding layered or unfamiliar logic
• AI stuck in a loop — Claude keeps repeating the same mistake or can't break out of a pattern

When Not to Use It

• Simple questions or text formatting tasks (default medium is enough)
• Frequent small tasks (using ultrathink every turn will spike your token costs fast)
• Subagent batch tasks (official recommendation is low effort for Subagents to save cost)

FAQ

What's the difference between ultrathink and just saying "think deeply"?

ultrathink is a keyword with a direct system mapping — Claude Code parses it as a high effort trigger. "Think deeply" is natural language that the model interprets semantically, which may or may not produce the same system behavior. In practice, ultrathink is more reliable and consistent.

Do I need to type ultrathink every turn?

Only on turns where you genuinely need deep reasoning. For most everyday tasks, the default medium effort is sufficient. If you need high effort throughout an entire session, use /model to set it permanently instead of typing ultrathink each time.

How many extra tokens does ultrathink use?

There's no fixed number — it depends on the complexity of your task. Compared to medium effort, high effort reasoning typically uses 2–5× more tokens. If you want to track actual usage, pair it with a usage monitoring tool.

References

• Claude Code Best Practices - Anthropic Official
• What is UltraThink in Claude Code - claudelog.com
• Claude Code Release Notes - claudelog.com
• Ultrathink is a Claude Code magic word - Hacker News