Model reports

ViBench

End-to-end vibe-coding benchmark by authors at Replit and Georgian AI Lab, scoring web applications from the user's perspective. Tasks are derived from anonymized Replit production traces, so the benchmark is first-party to one of the agents it measures. It reaches this registry through a customer testimonial in Anthropic's Claude Fable 5 / Mythos 5 launch post, where it is described qualitatively and never scored in the comparison table. Distinct from VBench, an unrelated video-generation benchmark.