# MirrorCode (Epoch AI and METR) (MirrorCode)

> MirrorCode, built by Epoch AI with METR, asks AI agents to reimplement entire command-line programs from scratch without their source code, matching the original's output exactly on end-to-end and held-out tests.

Canonical page: https://benchlm.ai/benchmarks/mirrorcode

- Category: [Coding](/coding)
- Last updated: November 13, 2026

## About MirrorCode

- Year: 2026
- Tasks: 30 reimplementation tasks (15 programs in two languages), three attempts each
- Format: Agentic black-box reimplementation, exact-output end-to-end tests
- Difficulty: Days-long software engineering without source code
- Paper: [MirrorCode (Epoch AI)](https://epoch.ai/benchmarks/mirrorcode)

Agents get execute-only access to the original program, its documentation and example tests, but not its source or the web. The 25 target programs cover Unix utilities, serialization and query tools, bioinformatics, interpreters, static analysis, cryptography and compression. Leaderboard scores use the MirrorCode (ML, +Private, 2L) setup: 15 medium and large programs, each in two implementation languages (generally Go and Ada), 30 tasks, three attempts each, with up to 10 billion tokens and seven days per attempt. A task counts as solved only if 100% of tests pass, including held-out tests averaging 34% of the suite, scored in a separate sandbox without the reference binary.

MirrorCode is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (38 models)

| Rank | Model | Creator | Score | Who ran it |
|------|-------|---------|-------|-----------|
| 1 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | Anthropic | 97.8% | Run by MirrorCode |
| 2 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 95.6% | Run by MirrorCode |
| 3 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 93.3% | Run by MirrorCode |
| 4 | [Claude Opus 5](/models/claude-opus-5) | Anthropic | 87.8% | Run by MirrorCode |
| 5 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | Anthropic | 78.9% | Run by MirrorCode |
| 6 | [GPT-6.1 Terra](/models/gpt-6-1-terra) | OpenAI | 76.7% | Run by MirrorCode |
| 7 | [Claude Fable 5](/models/claude-fable) | Anthropic | 70.0% | Run by MirrorCode |
| 8 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 65.6% | Run by MirrorCode |
| 9 | [DeepSeek V4.1 Flash 1017](/models/deepseek-v4-1-flash) | DeepSeek | 64.4% | Run by MirrorCode |
| 10 | [Claude Sonnet 5](/models/claude-sonnet-5) | Anthropic | 63.3% | Run by MirrorCode |
| 11 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | OpenAI | 61.1% | Run by MirrorCode |
| 12 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 60.0% | Run by MirrorCode |
| 13 | [Gemini 4 Argon](/models/gemini-4-argon) | Google | 58.9% | Run by MirrorCode |
| 14 | [Kimi K3](/models/kimi-k3) | Moonshot AI | 54.4% | Run by MirrorCode |
| 15 | [Muse Spark 1.3](/models/muse-spark-1-3) | Meta | 54.4% | Run by MirrorCode |
| 16 | [Grok 4.7](/models/grok-4-7) | xAI | 54.4% | Run by MirrorCode |
| 17 | [Qwen3.8 Max](/models/qwen3-8-max) | Alibaba | 52.2% | Run by MirrorCode |
| 18 | [GPT-6 Sol](/models/gpt-6-sol) | OpenAI | 48.9% | Run by MirrorCode |
| 19 | [Muse Wonder](/models/muse-wonder) | Meta | 45.6% | Run by MirrorCode |
| 20 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | OpenAI | 41.1% | Run by MirrorCode |
| 21 | [Grok 4.6](/models/grok-4-6) | xAI | 41.1% | Run by MirrorCode |
| 22 | [Gemini 3.7 Flash](/models/gemini-3-7-flash) | Google | 41.1% | Run by MirrorCode |
| 23 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | Google | 38.9% | Run by MirrorCode |
| 24 | [GLM-5.2](/models/glm-5-2) | Z.AI | 36.7% | Run by MirrorCode |
| 25 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 33.3% | Run by MirrorCode |
| 26 | [DeepSeek V4 Pro 0813](/models/deepseek-v4-pro-0813) | DeepSeek | 33.3% | Run by MirrorCode |
| 27 | [Grok 4.5](/models/grok-4-5) | xAI | 32.2% | Run by MirrorCode |
| 28 | [GPT-5.2](/models/gpt-5-2) | OpenAI | 32.2% | Run by MirrorCode |
| 29 | [Claude Opus 4.7](/models/claude-opus-4-7) | Anthropic | 31.1% | Run by MirrorCode |
| 30 | [Mistral Large 4](/models/mistral-large-4) | Mistral | 27.8% | Run by MirrorCode |
| 31 | [GPT-6 Luna](/models/gpt-6-luna) | OpenAI | 26.7% | Run by MirrorCode |
| 32 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 23.3% | Run by MirrorCode |
| 33 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 23.3% | Run by MirrorCode |
| 34 | [Gemini 3.6 Flash](/models/gemini-3-6-flash) | Google | 23.3% | Run by MirrorCode |
| 35 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | OpenAI | 21.1% | Run by MirrorCode |
| 36 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 20.0% | Run by MirrorCode |
| 37 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | Google | 8.9% | Run by MirrorCode |
| 38 | [GPT-5 (high)](/models/gpt-5-high) | OpenAI | 4.4% | Run by MirrorCode |

## FAQ

### What does MirrorCode measure?

Long-horizon coding: rebuilding a whole command-line program from its behavior, documentation and examples, without its source, so the new program matches the original's output exactly.

### Which model leads MirrorCode?

Claude Opus 5.5 at high effort leads with 78.9%, followed by Claude Fable 5.1 (73.3%), Claude Fable 5 (63.9%) and Gemini 4 Argon (58.9%).

## Compare Top Models on MirrorCode

- [Claude Sonnet 5.5 vs Claude Opus 5.5](/compare/claude-opus-5-5-vs-claude-sonnet-5-5)
- [Claude Opus 5.5 vs Claude Fable 5.1](/compare/claude-fable-5-1-vs-claude-opus-5-5)
- [Claude Fable 5.1 vs Claude Opus 5](/compare/claude-fable-5-1-vs-claude-opus-5)
- [Claude Opus 5 vs Claude Haiku 5.5](/compare/claude-haiku-5-5-vs-claude-opus-5)