browser_use 发布 odysseys2 基准:100 个真人网页任务,gpt-6 astra 完成 47 个
browser_use 和 CMU 搞了个很狠的浏览器智能体基准,任务直接在真实网页上跑,最长 60 分钟一个。gpt-6 astra 才过 47/100,opus 5.5 只有 18,还有 5 个模型挂零。做智能体评测的都该看看任务设计。
browser_use 联合 CMU 推出浏览器智能体基准 odysseys2,包含 100 个实时网页任务,每个任务最长 60 分钟。任务包括玩 openguessr 地理猜谜、把乐高说明书转换成零件清单、在 5 个国家获取二手交易报价等需要多步操作的真实网页场景。
当前最好成绩是 gpt-6 astra 完成 47 个任务,opus 5.5 完成 18 个,另有 5 个模型成绩为 0。这说明现有模型在长时间、多步骤的真实网页操作上仍有很大差距,单步能力测试的高分并不能代表实际浏览任务的可靠性。
对做智能体产品的开发者来说,odysseys2 提供了一套比静态截图任务更接近真实使用的评测标准,可以用来衡量自己产品在长时程浏览器自动化上的能力。
this is what the hardest browser agent benchmark looks like 👀 odysseys2: 100 live-web tasks, up to 60 min each. play openguessr, turn a lego manual into a parts list, get trade-in quotes in 5 countries. gpt-6 astra finishes 47. opus 5.5: 18. five models: 0. built with cmu ↓
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include odysseys2, browser_use, gpt-6 astra, opus 5.5.
Source attribution: curated from Browser Use (@browser_use). Cross-source corroboration: 11 independent sources have covered this development (including Lxfater, marktechpost, 掘金本周最热), confirming industry-wide signal strength.