AI rebuilt a 16,000-line program from scratch, a job estimated at 2–17 weeks for a person
Given only a working program and its tests, Claude Opus 4.6 recreated it almost perfectly. Months-long coding jobs are still too hard.
MirrorCode, a test from Epoch AI and METR, asks AI agents to recreate existing command-line programs exactly. They can run the original program and see its tests, but never its source code (the written instructions behind it). Hundreds to thousands of tests check the result.
In early results (April 2026), Claude Opus 4.6 passed 2,000 of 2,001 tests recreating gotree, a toolkit for biology data with about 16,000 lines of Go code. Researchers estimated it would take a human engineer 2–17 weeks. Older models scored far lower: 63% for Opus 4.5, 15% for Opus 4. A 61,000-line Java/Kotlin program was not solved.
In the full benchmark (June 2026, 25 programs), Claude Opus 4.7 solved gotree in 14 hours for $251. One run on a large task worked for 19 days without human help, at a cost of $2,600. Epoch says the hardest tasks would take a person months, and no tested model solved them.
What it means
When the target is clear, AI can now finish some coding jobs that would take a programmer weeks. Most real software work has no such complete description to copy, so this shows what is possible in ideal conditions, not in a typical job. The largest programs are still out of reach.
Caveats
- These tasks come with unusually complete instructions (a working program plus tests); most real software work doesn’t.
- The authors say memorisation of the open-source originals “cannot be ruled out”.
- The human time estimates are informal and wide.
- Per-model solve rates in the full benchmark come from secondary reporting.