Benchmark · Coding

MirrorCode

An agent rebuilds a whole program from its behaviour alone, without the source or the internet, and must match the original's output on held-out end-to-end tests. Epoch AI and METR built it; the leaderboard runs 15 medium and large programs in two languages, three attempts each, with up to 10 billion tokens and 7 days an attempt.

Score
% of tasks solved, higher is better
Run by
Epoch AI
Status
Current, newest result 17 November 2026
Updated
as models are released
Ceiling
The best result is within five points of a perfect score, so the test no longer tells the top models apart
Data
the published file, CC BY 4.0, read 20 November 2026
Method
epoch.ai/MirrorCode

Results

19 models we list, each at its best setting, best first.

#ModelScoreRangeWho ran itSettingDate
1Claude Sonnet 5.5 A97.8%Independenteffort max2026-10-05source
2Claude Opus 5.5 A95.6%Independenteffort max2026-10-30source
3Claude Fable 5.1 A93.3%Independenteffort max2026-09-09source
4Claude Haiku 5.5 A78.9%Independenteffort max2026-10-21source
5GPT-6.1 Terra AA76.7%Independenteffort max2026-10-30source
6GPT-6.2 Astra AA67.8%Independenteffort max2026-10-27source
7DeepSeek V4.1 Flash C64.4%Independenteffort max2026-10-21source
8GPT-6.1 Sol AA61.1%Independenteffort max2026-10-05source
9GPT-6 Astra A60.0%Independenteffort max2026-10-29source
10Gemini 4 Argon B58.9%Independenteffort high2026-10-14source
11Grok 4.7 BB54.4%Independenteffort xhigh2026-10-29source
11Muse Spark 1.3 D54.4%Independenteffort max2026-10-20source
11Kimi K3 B54.4%Independenteffort max2026-09-03source
14GPT-6 Sol A48.9%Independenteffort max2026-10-29source
15Muse Wonder D45.6%Independenteffort max2026-10-14source
16Kimi K3.1 Pro D43.3%Independenteffort max2026-11-17source
17Gemini 3.8 Flash AA38.9%Independenteffort high2026-09-02source
18GPT-6 Luna A26.7%Independenteffort max2026-10-21source
19Gemini 3.1 Pro (preview) AA8.9%Independenteffort high2026-08-10source
Every setting, 22 rows
ModelAs the leaderboard names itScoreWho ran itSettingDate
Claude Sonnet 5.5claude-sonnet-5-597.8%Independenteffort max2026-10-05source
Claude Opus 5.5claude-opus-5-595.6%Independenteffort max2026-10-30source
Claude Fable 5.1claude-fable-5-193.3%Independenteffort max2026-09-09source
Claude Haiku 5.5claude-haiku-5-578.9%Independenteffort max2026-10-21source
Claude Opus 5.5claude-opus-5-578.9%Independenteffort high2026-09-22source
GPT-6.1 Terragpt-6.1-terra76.7%Independenteffort max2026-10-30source
Claude Fable 5.1claude-fable-5-173.3%Independenteffort high2026-09-10source
GPT-6.2 Astragpt-6.2-astra67.8%Independenteffort max2026-10-27source
DeepSeek V4.1 Flashdeepseek-v4.1-flash64.4%Independenteffort max2026-10-21source
GPT-6.1 Solgpt-6.1-sol61.1%Independenteffort max2026-10-05source
GPT-6 Astragpt-6-astra60.0%Independenteffort max2026-10-29source
Gemini 4 Argongemini-4-argon58.9%Independenteffort high2026-10-14source
Grok 4.7grok-4.754.4%Independenteffort xhigh2026-10-29source
Muse Spark 1.3muse-spark-1.354.4%Independenteffort max2026-10-20source
Kimi K3kimi-k354.4%Independenteffort max2026-09-03source
GPT-6 Solgpt-6-sol48.9%Independenteffort max2026-10-29source
GPT-6 Astragpt-6-astra46.7%Independenteffort high2026-08-30source
Muse Wondermuse-wonder-1.045.6%Independenteffort max2026-10-14source
Kimi K3.1 Prokimi-k3.1-pro-together43.3%Independenteffort max2026-11-17source
Gemini 3.8 Flashgemini-3.8-flash38.9%Independenteffort high2026-09-02source
GPT-6 Lunagpt-6-luna26.7%Independenteffort max2026-10-21source
Gemini 3.1 Pro (preview)gemini-3.1-pro-preview8.9%Independenteffort high2026-08-10source

A rank is among the models we list with a result, not a place on the whole leaderboard, and a model offered by several hosts is shown once, on its own listing. These results don't move our grades. Data: Epoch AI, Capabilities and benchmarking (CC BY 4.0), CC BY 4.0.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.