RiskWebWorld: leaderboard

RiskWebWorld (2026) gives web GUI agents e-commerce risk-management tasks from a production risk-control pipeline, run on real, often uncooperative websites within 20 steps. Task success rate over all 1,513 tasks (Standard subset with and without SOP guidance plus the Challenge subset without SOPs), in percent; higher is better.

Metric: Task success rate (%) over all 1,513 tasks (Standard subset with and without SOP guidance plus the Challenge subset without SOPs), for web GUI agents in RiskWebWorld's Browser Use style loop (DOM tree, screenshot and history each step, 1920x1080 viewport, at most 20 steps, stopped after three consecutive failed actions), default model settings; exact match for exact fields, an LLM judge for free-text fields; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro49.1#145
2GPT-5.248.7#185
3Qwen 3 VL 235B A22B47.1#485
4Qwen 3.5 397B A17B42.4#264
5Kimi K2.532.6#266
6Claude Sonnet 4.532.4#261
7BU-30B32.4
8Qwen 3.5 35B A3B32#492
9Qwen 3 VL 30B A3B24.2#784
10GLM-4.6V6.3#710
11InternVL3.5-38B2.2#1021
12uitars-1.5-7B1.4#7905
13ShowUI-2B0
14GUI-G1-3B0
15UI-TARS-72B (RiskWebWorld checkpoint unspecified)0

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

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