Browser Use Bench v2.1 | 180 tasks | October 7, 2026

Claude Haiku 5.5on real browser tasks.

#11 of 19 at xhigh effort. DeepSeek V4.1 Flash max scores 1.4 points higher for 2.6× less.

Haiku ran with a 32k output-token cap per request; GPT-6 Luna could use 128k. Both models support 128k.

Score69.5

Cost / task46¢

Median time14 min

#	Model	Score	95% range	Cost / task	Median time
11	Claude Haiku 5.5 xhigh	69.5	65–73	46¢	14 min
14	Claude Haiku 5.5 max	65.3	61–70	$1.07	21 min
15	Claude Haiku 5.5 high	64.3	60–68	23¢	9 min
18	Claude Haiku 5.5 medium	53.4	50–57	9.2¢	5 min
19	Claude Haiku 5.5 low	45.6	42–50	2.7¢	2 min

180 long tasks on live websites, the same for every model, run in the open-source BrowserCode harness (Browser Use's own agent runs in ours). Mean weighted rubric score per task, 0-100. Estimated model cost per task in USD. Excludes search, judge, browser and runner costs. Runner and encrypted tasks: github.com/browser-use/benchmark.

Compare |full leaderboard

GPT-6.1 Sol79.6
GPT-6 Astra77.9
GPT-6 Sol74.4
GPT-5.6 Luna69.6
GPT-6 Luna69.0
Opus 5.575.0
Opus 566.8
Sonnet 5.572.7
Fable 5.172.0
Grok 4.773.3
DeepSeek V4.1 Flash70.9
GLM 5.3 FlashX59.7

Run your tasks in the cloud