Definition
Reasoning token optimization reduces internal chain-of-thought length in reasoning models without sacrificing task accuracy — cutting inference cost especially in multi-turn agent loops.
Key Points
- Reasoning models may spend >90% of tokens on internal thinking (ember-1 launch blog)
- Multi-turn agents replay prior reasoning each turn — quadratic context/cost growth
- ember-1: 71.3% reasoning token cut, 39% total token cut in one customer A/B
- Same API pricing × fewer tokens = direct cost savings