StressWeb: leaderboard

StressWeb (2026) is a diagnostic stress test for web agents: generated websites with clean reference workflows are re-run under controlled perturbations, and deterministic scripts check task checkpoints on the recorded page states. This board is the checkpoint pass rate in percent averaged over the clean environment and the six perturbed ones; higher is better.

Metric: Checkpoint pass rate (%): share of predefined task checkpoints (milestones such as reaching a target page or completing required fields) that the agent's run passes, averaged over the clean environment and the six perturbed environments, on 149 tasks over 10 generated websites, a multimodal agent driving a Playwright controller from screenshots, the page DOM and its action history (100-step budget), checkpoints verified by task-specific scripts on the recorded DOM states; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1O4 Mini50.1#368
2Gemini 2.5 Pro48.1#276
3Claude Opus 4.543.7#160
4Claude Sonnet 4.543.6#261
5GPT-5.241.7#185
6Qwen 3 VL 235B A22B (Thinking)40.3#459
7GLM-4.6V35.2#710
8Claude Haiku 4.534.6#583

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

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