Benchmarks · reasoning
Opus Magnum Bench
Uses each model variant's best pooled human-normalized reward across the full 36-puzzle campaign. Score: mean best human-normalized puzzle reward. Collected by Jake Boggs' benchmark dashboard.
20 models · 21 results · data as of 2026-10-29 · opusmagnumbench.com ↗
Leaderboard
GPT-5.6 Sol xhigh
All results
21 rows| # | Model | Effort | Opus Magnum Bench |
|---|---|---|---|
| 1 | GPT-5.6 Sol | xhigh | 70.9% |
| 2 | Fable 5 | max | 60.2% |
| 3 | GPT-5.5 | xhigh | 53.9% |
| 4 | GPT-5.6 Terra | xhigh | 47.7% |
| 5 | GPT-5.5 | medium | 37.9% |
| 6 | Gemini 3.1 Pro | high | 31.1% |
| 7 | Sonnet 5 | max | 26.5% |
| 8 | GLM 5.2 | high | 25.4% |
| 9 | Gemini 3.5 Flash | high | 24.5% |
| 10 | Opus 4.8 | max | 22.8% |
| 11 | Grok 4.5 | high | 18.2% |
| 12 | GPT-5.6 Luna | xhigh | 17.8% |
| 13 | Gemini 3 Flash | high | 15.4% |
| 14 | DeepSeek V4 Pro | max | 13.4% |
| 15 | Sonnet 4.6 | max | 12.4% |
| 16 | Kimi K2.7 Code | max | 10.9% |
| 17 | GLM 5.1 | max | 10.8% |
| 18 | GPT-5.4 Mini | xhigh | 9.2% |
| 19 | GPT-5.4 Nano | xhigh | 6.5% |
| 20 | Kimi K2.6 | max | 5.5% |
| 21 | Grok 4.3 | high | 0.0% |