Key findings
- Claude Opus 5.5 by Anthropic had the highest verification score, 85.0% (68/80); the median of 72 scored models was 62.5%.
- Among models at or above the median score, the cheapest correct answer came from GLM 5.3 Flash at $0.00022; the most expensive, Grok 4.7, cost 145× as much per correct answer.
- Reasoning tokens made up 86% of the output of the 58 models that think before answering. Their median cost per question was $0.0027, against $0.00045 for the 14 that don't.
- Nemotron 3.5 Super VL was the fastest model at a median 715 tokens per second.
