Benchmark
Whether a model can build a complete, working web application from a written product specification, scored by how many of the finished app's user-interface tests pass.
| Model | Score (%) | Source | Evaluated |
|---|---|---|---|
| Claude Opus 5.5single agent, medium effort | 91.5 | vals.ai IndependentVerified | — Source date: October 9, 2026 |
| Claude Opus 5.5single agent, maximum effort | 89.8 | vals.ai IndependentVerified | — Source date: October 9, 2026 |
| GPT-6 Solsingle agent, maximum effort | 89 | vals.ai IndependentVerified | — Source date: October 9, 2026 |
| GPT-6 Solsingle agent, medium effort | 77.6 | vals.ai IndependentVerified | — Source date: October 9, 2026 |
Fifty applications are built from product specs, deployed with Docker Compose, and then driven by browser agents that click through each app's workflows. The score is the share of UI tests passed. Vals AI runs each model both as a single agent and as a lead agent delegating to subagents, at medium and maximum reasoning effort.
Each configuration was run once, so single comparisons carry wide uncertainty and most differences Vals AI measured were not statistically significant. Browser-agent grading can miss defects a human reviewer would catch, and passing a workflow test is not the same as the application being correct, secure or maintainable.
Agent teams cost up to 5.1 times as much as one agent, and beat it in only one of four comparisons
On Vals AI's Vibe Code Bench, a lead agent delegating to subagents beat a single agent once out of four — and simply raising the single agent's reasoning effort did more for less.
Privacy preferences
We use necessary cookies to run this site. With your consent we also use advertising cookies. You can change your choice at any time. Cookie Policy · Privacy Policy
