UltraThink is Dead. Long Live Extended Thinking. — Decode Claude

The page says ultrathink no longer matters in Claude Code because extended thinking is now enabled by default for supported models. It keeps the default budget at 31,999 tokens, and says 64K-output models can be pushed to 63,999 tokens with MAX_THINKING_TOKENS=63999.

It also says the old keyword was deprecated on January 16, 2026, that extended thinking can be disabled with MAX_THINKING_TOKENS=0 or alwaysThinkingEnabled: false, and that more thinking is useful mainly for harder tasks where extra tokens and latency are worth it.

Reading notes

#

• ultrathink is deprecated and no longer changes Claude Code behavior
• supported models now get extended thinking automatically
• the default thinking budget is 31,999 tokens
• 64K-output models can use 63,999 thinking tokens with MAX_THINKING_TOKENS=63999
• the old behavior can be disabled with MAX_THINKING_TOKENS=0 or alwaysThinkingEnabled: false
• the page says more thinking fits complex system design, multi-file refactors, performance work, and other tasks where error cost is high
• the appendix links thinking tokens to test-time compute, chain-of-thought, scratchpads, and the idea that intermediate tokens support multi-step reasoning