Ultrathink is deprecated & How to enable 2x thinking tokens in Claude Code
4 months ago
- #AI
- #Extended Thinking
- #Machine Learning
- The 'ultrathink' keyword, which previously unlocked Claude's maximum reasoning power, is now deprecated.
- Extended thinking is now automatically enabled for supported models with a default budget of 31,999 tokens.
- A hidden trick allows users to get 63,999 thinking tokens on 64K output models by setting 'MAX_THINKING_TOKENS=63999'.
- Extended thinking is beneficial for complex tasks like system design and performance optimization but may not be necessary for simpler tasks.
- Users can disable extended thinking by setting 'MAX_THINKING_TOKENS=0' or 'alwaysThinkingEnabled: false' in settings.
- Thinking tokens expand the computational capabilities of transformers by allowing intermediate reasoning steps.
- Research shows that test-time compute (thinking tokens) can outperform model scaling in certain scenarios.
- Major AI labs like OpenAI, Anthropic, and Gemini have integrated thinking tokens into their flagship models.
- The trade-off of using more thinking tokens includes higher latency, cost, and diminishing returns for simple tasks.