WEBMCP.COM

WindTunnel. An open benchmark for WebMCP.

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

Tasks solved
48–49/49
tasks solved across WebMCP configurations
+8.9% median tasks solved vs other methods (45 of 49)
Faster
3.9×
median task completion time
vs the median configuration using other methods
Less cost
9×
median cost per task
vs the median configuration using other methods
Final score
+15–41%
higher final score
across WebMCP setups, vs the median of other methods

Leaderboard

Rank Configuration Interface Final score Tasks solved Attempt success Median costper task Median tokensper task Median timeper task
1Haiku 5.5🏆WebMCP98.349/49147/147 (100.0%)$0.0002515,1533.7s
2DeepSeek V4.1 Flash-1017WebMCP97.049/49147/147 (100.0%)$0.0005394,1543.3s
3GPT-6 LunaWebMCP95.548/49144/147 (98.0%)$0.0001942,5255.0s
4Muse EmberNewWebMCP95.148/49144/147 (98.0%)$0.0004254,6143.9s
5Jev + Mercury 2.5WebMCP92.849/49141/147 (95.9%)$0.0010649,7933.2s
6Sonnet 5.5WebMCP90.049/49147/147 (100.0%)$0.0035305,0464.2s
7GPT-6.1 SolWebMCP85.449/49147/147 (100.0%)$0.0034322,5028.0s
8Gemini 3.8 FlashWebMCP85.049/49147/147 (100.0%)$0.0036383,7998.3s
9GPT-6 AstraWebMCP82.249/49146/147 (99.3%)$0.0171002,5756.3s
10Muse WonderWebMCP80.649/49146/147 (99.3%)$0.0155495,2368.2s
11Claude Opus 5WebMCP80.049/49147/147 (100.0%)$0.0140134,7709.8s
12GPT-6.1 SolCode execution75.449/49147/147 (100.0%)$0.01900110,50316.5s
13GPT-6 LunaComputer use72.641/49128/147 (87.1%)$0.00211221,17719.4s
14Haiku 5.5Computer use72.541/49123/147 (83.7%)$0.00169837,60116.0s
15GPT-6 AstraCode execution70.349/49147/147 (100.0%)$0.11867610,98216.4s
16Sonnet 5.5Computer use69.847/49140/147 (95.2%)$0.03164137,33919.7s
17GPT-6.1 SolComputer use66.345/49134/147 (91.2%)$0.03983920,79720.8s
18Jev + Mercury 2.5DOM (ultrafast)63.425/4976/147 (51.7%)$0.00077813,8925.4s
19GPT-6 AstraComputer use61.545/49135/147 (91.8%)$0.26078020,56020.8s
20Claude Opus 5Computer use56.445/49134/147 (91.2%)$0.13885747,14150.4s

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.

Methodology

We operate eight real websites across a wide range of tasks — 49 tasks, three attempts each. There are four 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.

CODE EXECUTION

Browser code

The model writes code that the agent executes to read and drive the page.

Download the current results CSV. The methodology, code, task definitions, and earlier run transcripts are on github.com/nekuda-ai/WindTunnel.

Changelog

v1.5

Nov 11, 2026 Latest
  • Added Muse Ember via Meta with WebMCP (48/49 tasks). The board now has 20 configurations and 2,940 attempts.

v1.4

Oct 18, 2026
  • Added DeepSeek V4.1 Flash-1017 with WebMCP (49/49 tasks). The board now has 19 configurations and 2,793 attempts.

v1.3

Oct 15, 2026
  • Added Sonnet 5.5, Haiku 5.5, GPT-6 Luna, GPT-6.1 Sol, Gemini 3.8 Flash, and Muse Wonder. The board now has 18 configurations.
  • Retired Sonnet 5, GPT-5.6, and Gemini 3.6 from the current board. Their earlier results remain available.
  • Fixed one Sonnet grading check so correct answers count. Older published results are unchanged.
  • Held Opus 5.5 while we review provider refusals. Opus 5 stays for now.

v1.2

Sep 18, 2026
  • Added Jev + Mercury 2.5 with WebMCP and DOM controls without WebMCP.
  • Jev + Mercury 2.5 with WebMCP leads the final score. The board now includes 21 configurations.

v1.1

Sep 6, 2026
  • Added the complete_checkout WebMCP tool, fixing the checkout task on the online store.
  • Re-measured all eight WebMCP configurations. Checkout improved from 0/3 → 3/3 for each.
  • Added GPT-6 Astra benchmarks for WebMCP, screenshots, and code execution.

v1.0

Aug 20, 2026
  • Initial benchmark release: 16 configurations, 2,352 attempts.

View full changelog →