Vibe Code Bench 1-100 tests coding agents on sequential app changes

Vals AI built Vibe Code Bench 1-100 to test what happens after an AI coding agent gets an app working. Its earlier Vibe Code Bench asks a model to build a working application from scratch; this one hands the model a finished, working app and then asks it to keep changing it. According to the benchmark page, each task starts from a working Vibe Code Bench v1.1 application and delivers up to ten product requests one at a time. The launch video describes these as plain-English requests to add or revert features, fix bugs or migrate a database.

The scoring is strict about what already worked. Each step runs tests for the new request plus selected regression tests, which re-check earlier behavior that must still hold. A workflow passes only if every step succeeds, and breaking a feature from the base app or from an earlier step counts as a failure. That is why the scores are low. The video says even the best model could barely solve a third of the tasks, and the page, as fetched on October 22, 2026, lists Wonder Full first at 33.98% as of October 14. Breaking existing work is a large share of the problem: among graded failures at the first step where a run stopped, the page reports that 13.94% involved only regression checks and 25.32% involved both regression and new-request checks, so close to four in ten included something that had previously worked.

This explanation draws on the opening announcement post, an automatic transcript of its video (the visuals were not analyzed) and the benchmark page. Any later posts in the thread were not collected.