UXBench - Human-Rated Lift: leaderboard
UXBench (2026) evaluates LLMs as interaction-grounded UX judges: the model explores a runnable web fixture, writes a UX report, and a fixed repair agent edits the page from that report. This board is the blind human-rated lift: six UX and front-end experts rate the repaired and unrepaired versions of 41 fixtures on a 1 to 5 rubric, and the score is the mean paired improvement in rubric points. Higher is better.
Metric: Mean repair lift (rubric points on the 1-5 scale, signed): blind ratings by six UX and front-end experts of the fixture repaired from the model's UX report minus the unrepaired baseline, paired by site over the 41 fixtures and averaged over seven rubric dimensions; the model explores each fixture through a coverage-gated browser trajectory and a fixed repair agent applies its report; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 0.35 | #123 |
| 2 | Claude Sonnet 4.6 | 0.32 | #161 |
| 3 | GPT-5.4 Mini | 0.26 | #389 |
| 4 | GLM-5.1 | 0.25 | #179 |
| 5 | Gemini 3.1 Pro (Preview) | 0.21 | #94 |
| 6 | Kimi K2.5 | 0.21 | #256 |
| 7 | GPT-5.4 Nano | 0.2 | #635 |
| 8 | Qwen 3.6 Plus | 0.17 | #222 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=uxbench-human-rated-lift · How It Works · Data refreshed daily, snapshot 2026-11-13.