1 story
A new benchmark ran eight frontier models against real tasks from private production codebases, and the best one solved under 40%.
Withspecific · 4m