Vals AI's New Benchmark Shows No Model Clears 30% on Iterative Coding

Vals AI released Vibe Code Bench 1-100, a follow-on to its earlier coding benchmark that stops grading once an application first works. The new version starts from a complete, passing application and feeds a model up to ten sequential product-manager-style change requests, the kind of ongoing feature work a real engineer receives, then checks both whether the new request works and whether roughly a third of the test suite covering earlier features still passes. Vals AI's benchmark page reports that at launch, the top-scoring model, Claude Opus 5.5, reached only 30.36% overall, and close to 39% of graded failures involved breaking something that previously worked rather than only failing to finish the new request. That points to non-regressive iteration, not first-pass code generation, as the harder unsolved part of agentic coding.