Anthropic A/B Tests Dynamic Effort Levels and Token Budgets in Claude Code

Developers analyzing API responses from Anthropic flagship CLI tool, Claude Code, have uncovered active A/B testing of dynamic reasoning effort parameters.

The experimental feature dynamically throttles deep-thinking loops based on task complexity. Simple syntax edits and file moves receive lightweight instant responses, while complex refactoring requests engage high-depth reasoning chains.

Early user reports indicate the optimization dramatically reduces average response latency during routine coding sessions while preserving deep reasoning for intricate bugs.