WebForge-Bench - Level 2: leaderboard

WebForge-Bench (2026) tests browser agents on tasks set in automatically generated, reproducible websites with injected real-web noise, controlled for difficulty along seven dimensions. This board is task accuracy in percent on difficulty level 2 with screenshot and DOM input; higher is better.

Metric: Accuracy (%) on the difficulty level 2 tasks of the 934 WebForge-Bench tasks on generated, self-contained websites with injected real-web noise, up to 50 browser actions in a Chromium GUI, final-state answer checking; multimodal setting: screenshot plus DOM; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro82.1#145
2Claude Sonnet 4.574.7#261
3Kimi K2.573.8#266
4Gemini 3 Flash73.5#175
5GPT-5 Mini68.2#364
6GPT-5.265.9#185
7Qwen 3 VL 235B A22B50.3#485
8Gemini 2.5 Flash Lite33.5#964
9GPT-5 Nano25.9#906
10qwen3-omni-30B9.1#1608

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=webforge-bench-level-2 · How It Works · Data refreshed daily, snapshot 2026-11-21.