Percentage of cases where the model initially failed but successfully passed after one round of self-refinement.

Cost per million tokens (input+output average). Lower is better.

Weighted average of Pass/Refine rate, reasoning, STEM, coding, and utility benchmarks. Censorship is penalized.