Omni2Web - Direct Editing: leaderboard
Omni2Web (Han et al., 2026) asks multimodal models to edit a source webpage directly from 918 bilingual screen recordings of weakly referential web-editing requests, where speech, cursor movement and edit history identify the targets. Edit Fidelity Score in percent combines target grounding, edit fulfilment, preservation of unrelated content and revision handling, judged against step-level rubrics by Gemini 3.1 Pro; higher is better.
Metric: Edit Fidelity Score (%; Track A: 0.10 target grounding + 0.50 edit fulfillment + 0.20 no-damage + 0.20 revision handling per instance, averaged over 918 bilingual screen recordings of weakly referential webpage edit requests (speech plus cursor), 13,907 edit steps; the model edits the source HTML directly from the recording under its native input interface (omni: audio-visual video; vision-language: video or 1 fps frames plus manual transcript); rubric items judged by Gemini 3.1 Pro; unusable outputs score zero). Source: arxiv.org. Saturation forecast: Around March 2028. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 51.17 | #98 |
| 2 | Claude Opus 5 | 47.37 | #20 |
| 3 | GPT-5.6 Sol | 42.86 | #29 |
| 4 | Grok 4.6 | 41.45 | #44 |
| 5 | Kimi K3 | 40.63 | #53 |
| 6 | Qwen 3.8 Max | 38.6 | #55 |
| 7 | Seed 2.1 Pro | 34.48 | #133 |
| 8 | Gemini 3.1 Pro (Preview) | 34.08 | #91 |
| 9 | Qwen 3.5 Omni Plus | 29.56 | #237 |
| 10 | MiniMax-M3 | 29.49 | #265 |
| 11 | MiMo-V2.5 | 28.79 | #353 |
| 12 | Seed 2.0 Lite | 28.33 | #286 |
| 13 | Qwen 3.8 27B | 26.94 | #199 |
| 14 | Muse Spark 1.1 | 24.08 | #82 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=omni2web-direct-editing · How It Works · Data refreshed daily, snapshot 2026-11-12.