MirrorCode (Epoch AI and METR) (MirrorCode)
We show this table for reference; we do not rank on it.
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.
Benchmark score on MirrorCode — November 13, 2026
We compile the MirrorCode rows from benchmark-owner or independent runs. Claude Sonnet 5.5 leads the table at 97.8%, followed by Claude Opus 5.5 (95.6%) and Claude Fable 5.1 (93.3%). We do not use these results to rank models overall.
Claude Sonnet 5.5
Anthropic
Run by MirrorCodeClaude Opus 5.5
Anthropic
Run by MirrorCodeClaude Fable 5.1
Anthropic
Run by MirrorCode38 modelsCodingCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (38 models)
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5AnthropicRun by MirrorCode | 97.8% | Run by MirrorCode | Not reported | Closed |
| 2 | Claude Opus 5.5AnthropicRun by MirrorCode | 95.6% | Run by MirrorCode | Not reported | Closed |
| 3 | Claude Fable 5.1AnthropicRun by MirrorCode | 93.3% | Run by MirrorCode | Not reported | Closed |
| 4 | Claude Opus 5AnthropicRun by MirrorCode | 87.8% | Run by MirrorCode | Not reported | Closed |
| 5 | Claude Haiku 5.5AnthropicRun by MirrorCode | 78.9% | Run by MirrorCode | Not reported | Closed |
| 6 | GPT-6.1 TerraOpenAIRun by MirrorCode | 76.7% | Run by MirrorCode | Not reported | Closed |
| 7 | Claude Fable 5AnthropicRun by MirrorCode | 70.0% | Run by MirrorCode | Not reported | Closed |
| 8 | Claude Opus 4.8AnthropicRun by MirrorCode | 65.6% | Run by MirrorCode | Not reported | Closed |
| 9 | DeepSeek V4.1 Flash 1017DeepSeekRun by MirrorCode | 64.4% | Run by MirrorCode | Not reported | Open |
| 10 | Claude Sonnet 5AnthropicRun by MirrorCode | 63.3% | Run by MirrorCode | Not reported | Closed |
| 11 | GPT-6.1 SolOpenAIRun by MirrorCode | 61.1% | Run by MirrorCode | Not reported | Closed |
| 12 | GPT-6 AstraOpenAIRun by MirrorCode | 60.0% | Run by MirrorCode | Not reported | Closed |
| 13 | Gemini 4 ArgonGoogleRun by MirrorCode | 58.9% | Run by MirrorCode | Not reported | Closed |
| 14 | Kimi K3Moonshot AIRun by MirrorCode | 54.4% | Run by MirrorCode | Not reported | Open |
| 15 | Muse Spark 1.3MetaRun by MirrorCode | 54.4% | Run by MirrorCode | Not reported | Closed |
| 16 | Grok 4.7xAIRun by MirrorCode | 54.4% | Run by MirrorCode | Not reported | Closed |
| 17 | Qwen3.8 MaxAlibabaRun by MirrorCode | 52.2% | Run by MirrorCode | Not reported | Open |
| 18 | GPT-6 SolOpenAIRun by MirrorCode | 48.9% | Run by MirrorCode | Not reported | Closed |
| 19 | Muse WonderMetaRun by MirrorCode | 45.6% | Run by MirrorCode | Not reported | Closed |
| 20 | GPT-5.6 TerraOpenAIRun by MirrorCode | 41.1% | Run by MirrorCode | Not reported | Closed |
| 21 | Gemini 3.7 FlashGoogleRun by MirrorCode | 41.1% | Run by MirrorCode | Not reported | Closed |
| 22 | Grok 4.6xAIRun by MirrorCode | 41.1% | Run by MirrorCode | Not reported | Closed |
| 23 | Gemini 3.8 FlashGoogleRun by MirrorCode | 38.9% | Run by MirrorCode | Not reported | Closed |
| 24 | GLM-5.2Z.AIRun by MirrorCode | 36.7% | Run by MirrorCode | Not reported | Open |
| 25 | DeepSeek V4 Pro 0813DeepSeekRun by MirrorCode | 33.3% | Run by MirrorCode | Not reported | Open |
| 26 | Claude Opus 4.6AnthropicRun by MirrorCode | 33.3% | Run by MirrorCode | Not reported | Closed |
| 27 | GPT-5.2OpenAIRun by MirrorCode | 32.2% | Run by MirrorCode | Not reported | Closed |
| 28 | Grok 4.5xAIRun by MirrorCode | 32.2% | Run by MirrorCode | Not reported | Closed |
| 29 | Claude Opus 4.7AnthropicRun by MirrorCode | 31.1% | Run by MirrorCode | Not reported | Closed |
| 30 | Mistral Large 4MistralRun by MirrorCode | 27.8% | Run by MirrorCode | Not reported | Open |
| 31 | GPT-6 LunaOpenAIRun by MirrorCode | 26.7% | Run by MirrorCode | Not reported | Closed |
| 32 | GPT-5.5OpenAIRun by MirrorCode | 23.3% | Run by MirrorCode | Not reported | Closed |
| 33 | GPT-5.4OpenAIRun by MirrorCode | 23.3% | Run by MirrorCode | Not reported | Closed |
| 34 | Gemini 3.6 FlashGoogleRun by MirrorCode | 23.3% | Run by MirrorCode | Not reported | Closed |
| 35 | GPT-5.6 LunaOpenAIRun by MirrorCode | 21.1% | Run by MirrorCode | Not reported | Closed |
| 36 | GPT-5.6 SolOpenAIRun by MirrorCode | 20.0% | Run by MirrorCode | Not reported | Closed |
| 37 | Gemini 3.1 ProGoogleRun by MirrorCode | 8.9% | Run by MirrorCode | Not reported | Closed |
| 38 | GPT-5 (high)OpenAIRun by MirrorCode | 4.4% | Run by MirrorCode | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
How to read this leaderboard
Higher is better. Claude Opus 5.5 at high effort leads at 78.9%. Fifteen models have been run, so the board is sparse compared with Epoch's puzzle benchmarks.
Operator receipt: 38 sourced rows are currently displayable on this page; the leading published row is Claude Sonnet 5.5 at 97.8%.
Honest limit: Each attempt is very expensive, so few models are evaluated and some at lower effort than their maximum. The score reflects Epoch's agent scaffold and budget. BenchLM shows MirrorCode as display-only; it does not enter overall or category rankings.
Among the reported MirrorCode rows, Claude Sonnet 5.5 is first at 97.8%. The third row is 4.5 points behind. The broader top-10 range is 34.5 points, so the table still separates the published systems.
38 models have been evaluated on MirrorCode. The benchmark falls in the Coding category. We keep these results separate from the weighted global scoring system, so they remain source-specific evidence. MirrorCode is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
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
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.
Freshness and provenance
Version
MirrorCode 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
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%).
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.