Benchmarks
326 sources · 1,256 models × 2,956 benchmarks · 154,484 scoresShare of Co-Design Bench's 273 tasks in which every hidden acceptance criterion is judged PASS after the verifier's refutation pass, so the whole requested revision is delivered.
547th strongest of 779 fitted benchmarks
Signal-to-noise +3.91 for the Co-Design Bench family
Another benchmark represents this family in the fit.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6 Astra | not reported · Claude Code | 26.4 | |
| 2 | Claude Opus 5 | not reported · Claude Code | 18.3 | |
| 3 | Qwen3.8 Max (0902) | not reported · Claude Code | 16.5 | |
| 4 | Kimi K3 | not reported · Claude Code | 15.8 | |
| 5 | Qwen3.8 27B | not reported · Claude Code | 15.0 | |
| 6 | GPT-5.6 Sol | not reported · Claude Code | 14.3 | |
| 7 | GLM-5.3 Flash | not reported · Claude Code | 10.6 | |
| 8 | Opus 4.8 | not reported · Claude Code | 10.3 | |
| 9 | GLM-5.2 | not reported · Claude Code | 9.5 | |
| 10 | Grok 4.6 | not reported · Claude Code | 6.6 |