ProductsSelected

Browser Use 优化缓存命中后任务成本降至 $0.45,发布 1.0.4 版本

Why we picked it

Browser Use 自己测出缓存命中率只有 78%,改了 harness 后到 96%,每任务从 $0.67 降到 $0.45,分数没掉。他们的 v2.1 基准还显示 browser use 1.0 agent 得 77.5,比 codex 里最好的浏览器工具 harness 高出约 4 分。跑浏览器 agent 的开发者可以直接升到 1.0.4 省钱省时间。

Browser Use 团队发现自身评测中缓存命中率只有 78%,随后修复了评测框架,命中率提升到 96%。这一改动让每个任务成本从 $0.67 降到 $0.45(便宜 33%),单任务耗时从 5.0 分钟缩短到 4.6 分钟,而基准分数保持不变。修复后的版本以 browser-use==1.0.4 发布,可通过 uv 安装。

Body

这次优化源于 Browser Use Benchmark v2.1 的评测。该基准包含 200 个超难长程任务,任务主要来自与 CMU 合作的 Odysseys v2 项目。评测控制了变量:使用同一个模型 gpt 6.1 terra medium、同样的任务和同一个 judge,只改变 harness。结果显示,6 个浏览器工具在 codex 内运行(frontierharness 上的最佳 harness)得分在 62.4 到 73.5 之间,而 browser use 1.0 agent 拿到 77.5。

对跑浏览器 agent 评测或生产的团队来说,这说明 harness 层的 prompt 缓存设计直接影响成本和速度,同分数下三分之一的开销差距完全可以通过工程手段消除。

Original · Browser Use

the last eval made us realize our cache hit rate was kinda low (78%), so we fixed the harness now it's 96%. 33% cheaper ($0.67 -> $0.45 per task) and faster (5.0 -> 4.6 min), same score uv add "browser-use==1.0.4" Browser Use @browser_use We measured 7 browser harnesses on 200 very hard tasks same model (gpt 6.1 terra medium), same tasks, same judge. only the harness changes > Browser Use Benchmark v2.1: 200 super hard long horizon tasks, mostly from our Odysseys v2 collab with CMU + a few extra on top > 6 browser tools inside codex (top harness on frontierharness): 62.4 to 73.5 > browser use 1.0 agent: 77.5 🔗 View Quoted Tweet📊 6 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Products" 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 Browser Use, cache optimization, coding assistants.

Source attribution: curated from Browser Use (@browser_use).