6 ms·

curious about methodology. ive seen them post results for claude/codex when they only ran over benchmarks 3 times per model...