WindTunnel. An open benchmark for WebMCP.

Name: WindTunnel: WebMCP benchmark runs
Creator: nekuda
License: https://www.apache.org/licenses/LICENSE-2.0

WindTunnel compares WebMCP with other ways browser agents interact with websites.

nekuda-ai/WindTunnel Star us nekuda/windtunnel

Tasks solved

48/49

solved by every WebMCP configuration

+11.6% vs the median screen-driving agent (43 of 49)

Faster

3–5×

task completion

7.8s vs 28.1s median per task

Less cost

4–23×

per task

0.6 cents vs 5.5 cents median per task

Final score

+39%

higher median final score

91.2 vs 65.8 for screen-driving agents

Leaderboard

Rank	Configuration	Interface	Final score	Tasks solved	Attempt success	Median costper task	Median tokensper task	Median timeper task
1	GPT-5.6 Luna · native	WebMCP	98.4	48/49	143/147 (97.3%)	$0.002	2,596	5.7s
2	Gemini 3.6 Flash · native	WebMCP	94.0	48/49	144/147 (98.0%)	$0.004	4,453	7.2s
3	Gemini 3.6 Flash · Stagehand v4	WebMCP	92.7	48/49	143/147 (97.3%)	$0.004	4,371	8.0s
4	Sonnet 5 · native	WebMCP	91.2	48/49	144/147 (98.0%)	$0.009	5,172	6.8s
5	Sonnet 5 · Stagehand v4	WebMCP	89.5	48/49	144/147 (98.0%)	$0.010	5,161	8.1s
6	GPT-5.6 SOL · native	WebMCP	86.3	48/49	142/147 (96.6%)	$0.012	2,573	9.3s
7	Claude Opus 5 · native	WebMCP	86.0	48/49	144/147 (98.0%)	$0.014	4,770	9.8s
8	GPT-5.6 Luna	Computer use	75.2	45/49	134/147 (91.2%)	$0.017	20,914	18.3s
9	GPT-5.6 Luna	DOM + vision	70.0	43/49	130/147 (88.4%)	$0.033	29,561	19.8s
10	GPT-5.6 Luna	Accessibility tree	69.8	40/49	119/147 (81.0%)	$0.020	18,517	16.0s
11	Gemini 3.6 Flash	Computer use	67.3	43/49	130/147 (88.4%)	$0.020	23,857	33.7s
12	GPT-5.6 SOL	Computer use	65.8	46/49	134/147 (91.2%)	$0.063	16,235	27.3s
13	Sonnet 5	DOM + vision	64.2	48/49	145/147 (98.6%)	$0.210	64,424	29.3s
14	Sonnet 5	Accessibility tree	62.7	42/49	128/147 (87.1%)	$0.038	10,762	37.5s
15	Sonnet 5	Computer use	57.8	39/49	119/147 (81.0%)	$0.070	57,701	31.7s
16	Claude Opus 5	Computer use	56.6	45/49	134/147 (91.2%)	$0.139	47,141	50.4s

Final score: The final score combines attempt success rate (60%), median cost per task (20%), and median agent time per task (20%). Cost and time are log-normalized (large differences are compressed so extreme values do not dominate); higher is better. Tokens are shown separately and are not scored twice.

Methodology

We operate eight real websites across a wide range of tasks — 49 tasks, three attempts each. There are three ways to operate the website:

WEBMCP

Tool calling

The website exposes direct actions for the agent to call.

SCREENSHOTS

Computer use

The agent reads rendered images of the page and acts by coordinate.

PAGE STRUCTURE

DOM

The agent reads the page's DOM and accessibility tree.

Everything is fully reproducible — the methodology, code, task definitions, and full run transcripts are on github.com/nekuda-ai/WindTunnel.