UltraThink Is Dead. Long Live Extended Thinking (decodeclaude.com)

Anthropic has officially deprecated the "ultrathink" keyword, a feature that previously unlocked extended reasoning capabilities in its Claude models. As of January 16, 2026, extended thinking is now automatically enabled for several supported models, including Opus 4.5 and Sonnet 4.5, providing a default thinking budget of 31,999 tokens without the need for special keywords. This change simplifies the user experience while maintaining consistency in performance across models. More significantly, users can leverage an undocumented trick to increase the thinking budget to 63,999 tokens on 64K output models, effectively doubling the reasoning capacity for complex tasks. The shift from ultrathink to a default thinking mode reflects a growing trend within the AI/ML community that emphasizes computational efficiency at inference time. Enhanced reasoning capabilities are crucial for addressing complex challenges like performance optimization and system design, where the cost of errors can outweigh the expenses associated with token usage. Research has shown that allowing models to engage in intermediate reasoning steps significantly improves their performance on complex tasks. Therefore, this update not only benefits individual users but also contributes to advancing the overall understanding of effective computational strategies in language models, influencing future developments in AI reasoning capabilities.