SWE-bench

Reporting on SWE-bench from Lazy Tensor.

Search this sectionA place, person or subject

Benchmarks

Maintainers merge 24 points fewer AI patches than SWE-bench says

Blind code reviews of 296 AI-generated pull requests by open-source maintainers reveal a 24 percentage point merge gap below SWE-bench Verified scores, alongside new MirrorCode agent benchmarks.

Oct 18, 2026By Lazy Tensor3 min read

Green ink illustration on white background showing irregular watercolor patch shapes crossed by a bold dark green X with an overlapping circle containing a checkmark

Coding Agents SWE-bench METR