Benchmarks · coding
MirrorCode (Epoch AI, with METR)
Reimplement a large existing program from its behavior alone. Built by Epoch AI with METR; score is the share of tests passed. Data: Epoch AI, CC BY.
15 models · 15 results · data as of 2026-10-29 · epoch.ai ↗
Leaderboard
Opus 5.5 max
All results
15 rows| # | Model | Effort | Score | Harness |
|---|---|---|---|---|
| 1 | Opus 5.5 | max | 77.4% | Epoch AI |
| 2 | Fable 5.1 | high | 73.3% | Epoch AI |
| 3 | Fable 5 | high | 63.9% | Epoch AI |
| 4 | Gemini 4 Argon | high | 58.9% | Epoch AI |
| 5 | GPT-6 Astra | high | 46.7% | Epoch AI |
| 6 | Gemini 3.7 Flash | high | 41.1% | Epoch AI |
| 7 | Grok 4.6 | xhigh | 41.1% | Epoch AI |
| 8 | Grok 4.5 | high | 32.2% | Epoch AI |
| 9 | Opus 4.7 | high | 31.1% | Epoch AI |
| 10 | GPT-6 Luna | max | 26.7% | Epoch AI |
| 11 | GPT-5.6 Sol | high | 20.0% | Epoch AI |
| 12 | GPT-5.4 | high | 15.6% | Epoch AI |
| 13 | GPT-5.5 | high | 10.0% | Epoch AI |
| 14 | Gemini 3.1 Pro | high | 8.9% | Epoch AI |
| 15 | GPT-5 | high | 4.4% | Epoch AI |