WebCoderBench: leaderboard
Tests generation of native-HTML web apps from 1,572 real user requirements collected by an industrial partner, scored on 24 rule-based and LLM-judged metrics across nine perspectives such as code, visual and content quality and requirement alignment. The overall score sums the metric z-scores weighted by survey-derived user preferences, in percent of a standard deviation relative to the evaluated pool, higher is better; Liu et al. (Peking University), 2026.
Metric: Overall score: preference-weighted sum of metric z-scores (% of a standard deviation). Source: arxiv.org. Saturation forecast: Around January 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (High) | 13.81 | #163 (GPT-5) |
| 2 | GLM-4.5 | 11.76 | #525 |
| 3 | MiniMax-M2 | 7.89 | #669 |
| 4 | DeepSeek V3.1 Terminus | 6.26 | #442 |
| 5 | DeepSeek R1 0528 | 2.41 | #425 |
| 6 | Gemini 2.5 Pro | 0.27 | #262 |
| 7 | GPT-5 Codex (High) | -0.03 | #240 (GPT-5 Codex) |
| 8 | Qwen 3 235B A22B Instruct | -0.04 | #680 |
| 9 | Qwen 3 Coder Plus | -2.31 | #712 |
| 10 | DeepSeek V3.1 (Thinking) | -3.73 | #500 (DeepSeek V3.1) |
| 11 | DeepSeek V3.1 | -3.88 | #500 |
| 12 | GPT-4o (2024-11-20) | -38.65 | #864 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=webcoderbench · How It Works · Data refreshed daily, snapshot 2026-11-12.