Updated daily at 05:00 UTC
Benchmark
MirrorCode
Measures: Long-horizon coding: an agent reimplements a whole command-line program in another language from its behaviour alone, without the source, and must match its output on held-out end-to-end tests.
- Maintainer
- Epoch AI and METR
- Introduced
- 2026
- Status
- Active
- Saturation
- Headroom
- Typical provenance
- Independent
- In scoring set
- Yes
Leaderboard
Every result we hold, whoever ran it. Only independent and third-party rows count towards a score.
| Model | Score | Provenance | Source | Evaluated |
|---|---|---|---|---|
| Gemini 3.1 Pro Google DeepMind · High effort | 8.9% | Independent | Epoch AI | 10 Aug 2026 |
| GPT-5.5 OpenAI · High effort | 10.0% | Independent | Epoch AI | 10 Aug 2026 |
| GPT-5.4 OpenAI · High effort | 15.6% | Independent | Epoch AI | 10 Aug 2026 |
| GPT-5.6 Sol OpenAI · High effort | 20.0% | Independent | Epoch AI | 10 Aug 2026 |
| Claude Opus 4.7 Anthropic · High effort | 31.1% | Independent | Epoch AI | 12 Aug 2026 |
| GPT-6 Astra OpenAI · High effort | 46.7% | Independent | Epoch AI | 30 Aug 2026 |
| Claude Fable 5 Anthropic · High effort | 63.9% | Independent | Epoch AI | 10 Aug 2026 |
| Claude Fable 5.1 Anthropic · High effort | 73.3% | Independent | Epoch AI | 10 Sep 2026 |
| Claude Opus 5.5 Anthropic · Max effort | 77.4% | Independent | Epoch AI | 22 Sep 2026 |
- Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks, CC BY 4.0.