Epoch AI, METR launch MirrorCode benchmark

// 2h agoBENCHMARK RESULT

Epoch AI, METR launch MirrorCode benchmark

Developed by Epoch AI and METR, MirrorCode evaluates the autonomous capabilities of AI coding agents over extended horizons by requiring them to rebuild entire programs from scratch. Testing models on 25 diverse tasks, the benchmark highlights extreme compute requirements, with some tasks running for up to 19 days and costing up to $2,600.

// ANALYSIS

MirrorCode shifts evaluation from simple code-generation tasks to complex architectural reconstruction, proving that while autonomous AI developers are becoming viable, their compute requirements are currently massive and economically impractical.

* Black-box reverse-engineering forces AI to demonstrate true functional understanding rather than code memorization.

* A $2,600 run cost over 19 days highlights the efficiency bottleneck of current long-horizon reasoning.

* Recreating multi-thousand line tools demonstrates that future SWE agents could automate legacy system migrations and ports.

// TAGS

agentbenchmarkingsoftware-engineeringepoch-aimetrmirrorcodelong-horizon-coding

DISCOVERED

2h ago

2026-06-29

PUBLISHED

2h ago

2026-06-29

RELEVANCE

9/ 10

AUTHOR

AI Revolution

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

POLICY54m ago

Age verification laws force identity attribution

Age verification regulations across the US, Europe, and Australia fundamentally serve as identity attribution systems that link digital accounts to real-world identities. The setup could lead to automated tracking of online speech, prompting warnings to resist verification or pay with privacy-focused methods like Monero.

OPEN SOURCE1h ago

PDFx bundles multiple documents into single PDF

PDFx is an open-source extension to the PDF standard that stores multiple files inside a single valid PDF using an embedded JSON manifest. Its companion desktop application displays the documents on a Figma-style 2D canvas for easy organization while maintaining compatibility with standard PDF readers.

BENCHMARK2h ago

Browser Use launches interactive LLM benchmark

Browser Use released a web development benchmark evaluating Claude Opus 4.7, GLM 5.2, GPT 5.5, Gemini 3.5 Flash, and Minimax M3 on 15 prompts from the public LLM Arena dataset. Utilizing the Browser Use Cloud API v4, each model generated fully interactive web applications and UI prototypes to evaluate real-world browser-based agent performance.