📝Ultrathink is a keyword in 📝Claude Code that, when included anywhere in a prompt, asks 📝Claude to reason more deeply on that single turn.
The keyword dates from 2025, when Claude Code mapped prompt words to fixed thinking budgets and "ultrathink" unlocked the largest, 31,999 tokens. Developer Simon Willison documented the mapping in April 2025 by reading the Claude Code source, which showed it was a client-side feature rather than a model capability.
Its behavior has since changed. 📝Anthropic removed the keyword triggers in January 2026, then reintroduced "ultrathink" in Claude Code v2.1.68 on March 4, 2026 to enable high effort for the next turn. On current models thinking is adaptive, and the standing depth of reasoning is set with effort levels through the /effort command; ultrathink acts as a one-turn override rather than a fixed token budget. Weaker phrases such as "think hard" no longer receive special handling.
Ultrathink suits architecture decisions, difficult debugging, and planning, where extra reasoning improves the answer. It adds latency and token cost, so it is wasted on lookups and routine edits. A separate keyword, "ultracode," triggers Claude Code's multi-agent workflow mode and is unrelated.