/ultrathink — Force deeper reasoning on hard problems

What it does

/ultrathink switches Claude into extended thinking mode for the next response. The model spends more tokens reasoning internally before it streams the answer back to you. You see a brief thinking-block summary, then the answer — same conversation, deeper pass.

Think of it as the "give it a minute" button. Default mode is fast but shallower; ultrathink is slower but more deliberate.

When to use it

• Architecture decisions — weighing two designs with non-obvious trade-offs.
• Concurrency bugs — race conditions, deadlocks, memory model questions.
• Algorithmic problems — picking the right data structure, complexity analysis.
• Code review of subtle PRs — security-adjacent diffs, refactors that touch invariants.

Skip it for: boilerplate, renames, "what does this function do" questions, looking up syntax.

Try it in your terminal

The simulator below shows the prompt a teammate posts. Type the command that triggers extended thinking and hit Enter.

Gotchas

• It only applies to the next turn. If you forget and type something else first, the slow mode is wasted.
• Token cost can go up 3–5x for that turn. Worth it for hard problems, burns budget on easy ones.
• Some hosted setups limit extended thinking on smaller models — check your model with /model before relying on it.

Read more

Slash commands reference →