Anthropic has deprecated the `ultrathink` keyword, which previously granted 31,999 thinking tokens for advanced reasoning. Extended thinking is now automatically enabled for supported models like Opus 4.5 and Sonnet 4.5, using the same 31,999 token budget by default. This removes the need for magic prompt words to unlock deeper model reasoning.

The shift reflects a broader industry move toward inference-time compute, where models think longer during generation rather than relying solely on size. This approach, formalized in chain-of-thought research, lets transformers solve complex serial problems by using intermediate tokens as a computational scratchpad, breaking past inherent architectural limits.

For 64K output models, a hidden setting unlocks double the default thinking budget. Setting the environment variable `MAX_THINKING_TOKENS=63999` provides 63,999 tokens—ideal for complex system design or multi-file refactors. Users can also disable thinking entirely via settings or environment variables, reverting to the old behavior.