All Benchmarks/Coding

MirrorCode

Higher is better

Made by Epoch AI with METR: the AI reimplements an existing program (Unix tools, compression, cryptography, interpreters, and more) from scratch without seeing its source code. It succeeds only if every output matches the original exactly, including on hidden tests, with up to seven days per attempt. Scores run from 0 to 100%. Higher is better.

Top score77.4%Claude Opus 5.5
Models tested9
Last updated2026.09.22
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    Anthropic
    Claude Opus 5.5
    2026.09.22 · max
    77.4%
  2. 2
    Anthropic
    Claude Fable 5.1
    2026.09.01 · high
    73.3%
  3. 3
    Anthropic
    Claude Fable 5
    2026.06.09 · high
    63.9%
  4. 4
    OpenAI
    GPT-6 Astra
    2026.09.04 · high
    46.7%
  5. 5
    Anthropic
    Claude Opus 4.7
    2026.04.16 · high
    31.1%
  6. 6
    OpenAI
    GPT-5.6 Sol
    2026.07.09 · high
    20.0%
  7. 7
    OpenAI
    GPT-5.4
    2026.03.05 · high
    15.6%
  8. 8
    OpenAI
    GPT-5.5
    2026.04.23 · high
    10.0%
  9. 9
    Google
    Gemini 3.1 Pro
    2026.02.19 · high
    8.9%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.