LongWebBench - Single-Image Input: leaderboard
LongWebBench (2026) evaluates vision-language models on generating long, multi-screen webpages from screenshots. This board gives each model one long screenshot of each of 490 real webpages and scores the structural fidelity of the generated page (0 to 10) with a GPT-4o judge on page scale, layout, section hierarchy, visual styling and information density, averaged over seven site categories; inputs the model or API rejects score 0. Higher is better.
Metric: Structural fidelity overall score (0-10): macro-average over seven site categories for full-page webpages generated from one long screenshot of each of 490 real webpages: a GPT-4o evaluator (gpt-4o-2024-11-20, deterministic) compares the rendered page with the reference on page scale, global layout, section hierarchy, visual styling (0.8 judge plus 0.2 DINO similarity) and information density; inputs the model or API rejects score 0; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 6.77 | #113 |
| 2 | GPT-5.2 | 6.5 | #176 |
| 3 | Gemini 3 Flash (Preview) | 6.36 | #144 |
| 4 | GLM-4.6V | 5.66 | #701 |
| 5 | Qwen 3 VL 235B A22B Instruct | 5.1 | #522 |
| 6 | Claude Opus 4.5 (20251101) | 3.84 | #132 |
| 7 | Claude Sonnet 4.5 (Thinking) | 3.81 | #250 (Claude Sonnet 4.5) |
| 8 | Seed-1.6 | 3.66 | #448 |
| 9 | Qwen 3 VL 8B Instruct | 3.47 | #981 |
| 10 | InternVL3-78B | 3.4 | #738 |
| 11 | GPT-4o | 3.23 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longwebbench-single-image-input · How It Works · Data refreshed daily, snapshot 2026-11-13.