Epoch AI and METR publish MirrorCode results for long-horizon coding tasks

Event Summary

Epoch AI and METR published MirrorCode results on long-horizon coding tasks that require models to reimplement programs without source-code or web access. One Claude Opus 4.7 run completed a 16,000-line target in 14 hours at a reported $251 inference cost.

Context & Narrative

MirrorCode uses end-to-end tests over 25 target programs. The results show that models can complete some substantial reimplementation tasks under the benchmark’s conditions. Epoch also reports a 56% headline score for Opus 4.7 and warns that target programs may have appeared in training data.

Key Findings

• Fact Grade B

  MirrorCode evaluates end-to-end program reimplementation without source-code or web access across 25 target programs.

  Sources [1][2]
• Impact Grade B

  Epoch reported that one Claude Opus 4.7 run reimplemented a 16,000-line target in 14 hours at a $251 inference cost.

  Sources [1]
• Limitation Grade B

  Epoch reports that pretraining exposure to open-source targets may inflate benchmark performance.

  Sources [1]

Impact Assessment

• Capability Leap +2 · Medium-term

  MirrorCode provides evidence that current models can complete some long-horizon program-reimplementation tasks under controlled conditions.

  Affected Groups: software engineers, AI-agent researchers, developer-tool builders

Consensus & Sources

• 1
  MirrorCode: What's the largest software project AI can complete on its own?

  URL: https://epoch.ai/MirrorCode

  Reference Evidence Citation logged Live source
• 2
  MirrorCode: Evidence that AI can already do some weeks-long coding tasks

  URL: https://metr.org/blog/2026-04-10-mirrorcode-preliminary-results/

  Reference Evidence Citation logged Live source
• 3
  Import AI 466: The bitter lesson for robotics, AIs complete week-long programming tasks

  URL: https://importai.substack.com/p/import-ai-466-the-bitter-lesson-for

  Reference Evidence Citation logged Live source