Research2026-08-03
Epoch AI, working with METR, published MirrorCode, a test that asks AI models to rebuild entire existing programs from scratch without seeing the original code, checked against hidden end-to-end tests. Claude Opus 4.7 rebuilt a 16,000-line bioinformatics toolkit in 14 hours for $251, work Epoch estimates would take a human engineer weeks. Epoch cautions that the models likely saw these open-source programs during training, which may inflate the scores.
What changed
Coding benchmarks mostly covered short tasks like bug fixes with $1-10 inference budgets.
What it unlocks
Measuring whether a model can rebuild a whole program from its behaviour alone, over days of unattended work.
- 25 target programs, 22 open-sourced
- gotree: 14 hours, $251
- largest task: $2,600, 19 days
Sources