Replit says most failed AI-built apps pass single-feature tests and break at interactions

Replit says about three-quarters of the failed builds from top AI models pass every single-feature check and break only where features meet.

The company gives four examples: a draft pull request that hides a merge conflict, comments that cannot attach to text inside table cells, an undo that erases a teammate's change, and a server owner without roles who cannot moderate.

The figure comes from Replit's X post on October 28, 2026. The post does not name the models tested or describe how the builds were evaluated.