Best model for browser agents.

Name: Browser Use Bench v2.1 model results
Creator: Browser Use

Browser Use Bench v2.1: 180 real tasks on live websites. Updated Oct 29, 2026.

Top score81.3GPT-6.1 Sol high$1.92 a task Best value79.0GPT-6.1 Terra high$1.26 a task · 35% less Our agent72.9Browser Use11¢ a task · 18× lessBest score at 11¢ or less

How we test.

Same tasks and judge for every model.

180 real tasks; the agent works through each on live websites; a judge checks a weighted rubric; every model gets a score from 0 to 100, plus cost and time per task.

Score vs. cost.

Up and to the left is better.

Model cost only. Lab models ran in the open-source BrowserCode harness, ours in Browser Use.

Leaderboard.

Best setting per model.

#	Model	Score	Cost / task	Minutes
	Browser UseBU fast	72.9	11¢	9.2
1	GPT-6.1 Solhigh	81.3	$1.92	14.3
2	GPT-6 Solhigh	79.4	$1.86	11.0
3	GPT-6.1 Terrahigh	79.0	$1.26	8.8
4	GPT-6 Astramedium	77.9	$10.77	8.7
5	Kimi K3.1 Prohigh	76.7	$1.89	8.8
6	Claude Opus 5.5high	75.0	$2.35	11.8
6	DeepSeek V4.1 Flash 1017max	75.0	22¢	16.8†
8	Grok 4.7high	73.3	$3.46	17.6
9	Claude Sonnet 5.5high	72.7	$1.53	9.4
10	Claude Fable 5.1high	72.0	$5.19	17.2
11	DeepSeek V4.1 Flashmax	70.9	18¢	17.5
12	GPT-5.6 Lunaxhigh	69.6	28¢	14.6
13	Claude Haiku 5.5xhigh	69.5	46¢	14.2
14	GPT-6 Lunamax	69.2	22¢	16.6
15	Claude Opus 5	66.8	$4.72	19.0
16	GLM 5.3 FlashXmax	59.7	55¢	14.1
17	MiniMax M3.1 Flashmedium	50.6	13¢	47.1

Show all 40 settings

#	Model	Score	Cost / task	Minutes
	Browser UseBU fast	72.9	11¢	9.2
1	GPT-6.1 Solhigh	81.3	$1.92	14.3
2	GPT-6.1 Solmedium	79.6	$1.30	7.9
3	GPT-6 Solhigh	79.4	$1.86	11.0
3	GPT-6 Solxhigh	79.4	$3.00	15.7
5	GPT-6.1 Terrahigh	79.0	$1.26	8.8
6	GPT-6.1 Solxhigh	78.5	$2.23	18.5
7	GPT-6 Astramedium	77.9	$10.77	8.7
8	Kimi K3.1 Prohigh	76.7	$1.89	8.8
9	DeepSeek V4.1 Flash 1017max	75.0	22¢	16.8†
9	Claude Opus 5.5high	75.0	$2.35	11.8
11	GPT-6.1 Sollow	74.7	75¢	4.6
12	GPT-6 Solmedium	74.4	$1.02	5.9
13	DeepSeek V4.1 Flash 1017low	74.0	12¢	11.6†
14	DeepSeek V4.1 Flash 1017high	73.6	15¢	12.6†
15	Grok 4.7high	73.3	$3.46	17.6
16	Claude Sonnet 5.5high	72.7	$1.53	9.4
17	Claude Fable 5.1high	72.0	$5.19	17.2
18	GPT-6 Solmax	71.5	$4.71	21.6
19	DeepSeek V4.1 Flashmax	70.9	18¢	17.5
20	GPT-5.6 Lunaxhigh	69.6	28¢	14.6
21	Claude Haiku 5.5xhigh	69.5	46¢	14.2
22	GPT-6 Lunamax	69.2	22¢	16.6
23	GPT-6 Lunaxhigh	69.0	16¢	11.3
24	GPT-6 Lunahigh	68.9	12¢	12.0
25	Claude Opus 5	66.8	$4.72	19.0
26	Claude Haiku 5.5max	65.3	$1.07	21.0
27	GPT-6 Lunamedium	64.3	7.7¢	7.8
27	Claude Haiku 5.5high	64.3	23¢	9.3
29	GPT-6 Solnone	63.5	$1.21	5.4
30	GPT-6 Sollow	63.3	44¢	3.5
31	GLM 5.3 FlashXmax	59.7	55¢	14.1
32	Claude Haiku 5.5medium	53.4	9.2¢	5.2
33	MiniMax M3.1 Flashmedium	50.6	13¢	47.1
34	MiniMax M3.1 Flashhigh	49.5	13¢	47.5
35	MiniMax M3.1 Flashlow	48.6	11¢	44.3
36	Claude Haiku 5.5low	45.6	2.7¢	2.4
37	GPT-6 Lunanone	37.6	9.1¢	3.7
38	GPT-6 Lunalow	36.2	1.3¢	1.4
39	GPT-5.6 Lunanone	31.0	4.0¢	2.1

† Ran on different CI runners, so its time is not comparable.

Tasks: 180 long tasks, live sites
Judge: GPT-5.6 Luna xhigh, weighted rubric
Cost: Model tokens only
Time: Median minutes per task

Runner, tasks and rubrics on GitHub

FAQ

Which LLM is best for browser agents?

GPT-6.1 Sol high: 81.3 of 100 at $1.92 a task, on 180 tasks on live websites. Next is GPT-6 Sol high at 79.4.

Which model is the best value?

GPT-6.1 Terra high: 79.0 at $1.26 a task, 2.3 points under the top for 35% less. It is the cheapest setting within 3 points of the top score.

How does Browser Use's own agent score?

72.9 at 11¢ a task, 18× less than the top score. No setting in the test scores higher at that price or below.

How is it scored?

An LLM judge (GPT-5.6 Luna xhigh) checks each task against a weighted rubric. Mean weighted rubric score per task, 0-100. Estimated model cost per task in USD. Excludes search, judge, browser and runner costs.

What about earlier benchmarks?

Internal Bench Hard and BU Bench V1 came before Browser Use Bench v2.1. Browser Use's own agent scored highest on both.

Start with $1 free.

One per person. Never expires.

Get API key