Browser Use Bench v2.1 | 180 tasks | October 17, 2026
Agent benchmarks.
Best LLM for browser agents (updated October 17, 2026)
GPT-6.1 Sol medium scores highest of the 14 lab models on Browser Use Bench v2.1: 79.6 out of 100 at $1.30 of model cost per task, on the same 180 live-website tasks. The cheapest lab setting within 10% of that score is DeepSeek V4.1 Flash 1017 low: 74.0, 5.6 points lower, at 12¢ a task, 11× less. The fastest of those timed on the same runners is GPT-6 Sol medium: 74.4, 5.2 points lower, with a median of 5.9 minutes a task against 7.9 for the leader.
Browser Use's own agent scores 72.9 at 11¢ of model cost a task, level with Claude Sonnet 5.5 high at 14× less. The lab setting closest to it on cost is DeepSeek V4.1 Flash 1017 low: 74.0 at 12¢ a task.
Updated
- OpenAI
- Anthropic
- xAI
- DeepSeek
- Z.ai
- Browser Use
Methodology
- Tasks
- 180 long tasks on live websites: the ones in the 200-task Browser Use Bench v2.1 set that every listed setting got a score on, so every model is scored on the same 180.
- Scoring
- The findings-v21 judge, running GPT-5.6 Luna xhigh, scores each task against its weighted findings rubric. Mean weighted rubric score per task, 0-100. The 95% range is a bootstrap over the per-task scores.
- Harness
- Lab models run in the open-source BrowserCode harness; Browser Use's own agent runs in the Browser Use harness. Some models are listed at more than one reasoning effort; each row is one effort.
- Cost
- Estimated model cost per task in USD. Excludes search, judge, browser and runner costs.
- Time
- Median minutes per task.
Runner, encrypted tasks and rubrics: github.com/browser-use/benchmark.
How this compares to other agent benchmarks
Browser Use Bench v2.1 runs long tasks on live websites, scores each against a weighted rubric, and reports model cost and time next to the score. Other public benchmarks measure different things:
- WebArena: self-hosted, reproducible websites in four domains (shopping, a forum, software development, content management), with each task checked for functional correctness.
- BrowseComp: OpenAI's 1,266 questions that take persistent browsing to answer, each with a short answer checked against a reference.
- OSWorld: 369 tasks on a real Ubuntu computer across web and desktop apps (plus 43 supplementary Windows tasks), each checked by an execution script. It tests the whole desktop, not just the browser.
- Online-Mind2Web: 300 tasks on 136 live websites, judged by humans and by an LLM judge the authors report agrees with humans about 85% of the time.
How we ran Online-Mind2Web with our own judge: our Online-Mind2Web write-up.
FAQ
- Which LLM scores highest on Browser Use Bench v2.1?
- Among the lab models, GPT-6.1 Sol medium, at 79.6 out of 100 (95% range 75.8 to 83.3). Next is GPT-6 Astra medium at 77.9.
- When was this leaderboard last updated?
- October 17, 2026. A model is added once we have run it on the same 180 tasks, its model cost is recorded and we publish it; each published model gets its own page.
- What does a task cost in model fees on Browser Use Bench v2.1?
- Across the lab-model settings in the table, from 2.7¢ (Claude Haiku 5.5 low, scoring 45.6) to $10.77 (GPT-6 Astra medium, scoring 77.9). The leader costs $1.30. Estimated model cost per task in USD. Excludes search, judge, browser and runner costs.
Each model
›Earlier benchmarks: Internal Bench Hard, OnlineMind2Web, BU Bench V1
Internal Bench Hard
Browser Use: 82% of tasks solved at 17¢ each. That is 20 points better than Opus 5, which costs 20× more per solved task.