Ultra-long-horizon software engineering.
FrontierSWE v2 asks agents to complete projects that take expert engineers days. Examples include porting git to Zig, writing a Lean 4 kernel type checker in Pascal, optimizing FFmpeg's libswscale, and training a speech decoder from MEG signals. Every task is scored continuously against published or measured anchors, so models earn partial credit for partial progress. It remains a significant unsaturated public coding benchmark.