# MirrorCode: AI Reimplements a 16k-Line Codebase — Weeks of Human Work in One Run

> Claude Opus 4.6 autonomously reimplemented a 16k-line bioinformatics toolkit from scratch — no source code, just an executable and tests.

- Org: Epoch AI / METR
- Date: 2026-04-10
- Category: benchmark, paper
- Source: https://epoch.ai/blog/mirrorcode-preliminary-results
- Canonical page: https://ai-tldr.dev/releases/mirrorcode-benchmark/

## In short

Epoch AI and METR released MirrorCode, a long-horizon coding benchmark where AI must reimplement real CLI programs from execute-only access. Claude Opus 4.6 reimplemented gotree (16,905 lines, 40+ commands) passing 99.95% of 2,001 tests — a task estimated at 2–17 weeks of human engineering.

## What it is

MirrorCode is a long-horizon software engineering benchmark from Epoch AI and METR. Each task gives an AI agent execute-only access to a real CLI program (no source) and a test suite, then asks it to reimplement the program from scratch. The benchmark spans Unix utilities, bioinformatics tools, cryptography, interpreters, and configuration languages — programs ranging from ~900 to 61,000 lines of code.

## How it works

The agent runs the reference program to observe its behavior, forms hypotheses about its implementation, writes code, and iterates against the test suite — all within a fixed token budget (up to 1 billion tokens). Performance is measured by exact output matching across hundreds to thousands of end-to-end test cases. Claude Opus 4.6 solved gotree (a 16,905-line Go bioinformatics toolkit with 40+ subcommands), passing 99.95% of 2,001 tests. Smaller programs like cal and choose were solved by earlier models; the 61k-line Pkl configuration language remains unsolved at ~35% pass rate.

## Why it matters

This is the first benchmark to put a concrete number on AI's reach into weeks-long engineering tasks. The gotree result demonstrates that a frontier model can reverse-engineer a complex, multi-command production tool entirely from behavioral observation. The token budget and test-pass metric give teams a replicable framework for measuring how far this capability extends as models improve.

## Who it's for

AI safety researchers tracking capability timelines; engineering teams evaluating AI for autonomous coding tasks

## Try it

https://epoch.ai/blog/mirrorcode-preliminary-results

## Links

- [Epoch AI Blog](https://epoch.ai/blog/mirrorcode-preliminary-results)
- [METR Blog](https://metr.org/blog/2026-04-10-mirrorcode-preliminary-results/)
- [Data Repo](https://github.com/epoch-research/MirrorCode-data)