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.
WindTunnel compares WebMCP with other ways browser agents interact with websites.
| Rank | Configuration | Interface | Final score | Tasks solved | Attempt success | Median costper task | Median tokensper task | Median timeper task |
|---|---|---|---|---|---|---|---|---|
| 1 | Haiku 5.5🏆 | WebMCP | 98.3 | 49/49 | 147/147 (100.0%) | $0.000251 | 5,153 | 3.7s |
| 2 | DeepSeek V4.1 Flash-1017 | WebMCP | 97.0 | 49/49 | 147/147 (100.0%) | $0.000539 | 4,154 | 3.3s |
| 3 | GPT-6 Luna | WebMCP | 95.5 | 48/49 | 144/147 (98.0%) | $0.000194 | 2,525 | 5.0s |
| 4 | Muse EmberNew | WebMCP | 95.1 | 48/49 | 144/147 (98.0%) | $0.000425 | 4,614 | 3.9s |
| 5 | Jev + Mercury 2.5 | WebMCP | 92.8 | 49/49 | 141/147 (95.9%) | $0.001064 | 9,793 | 3.2s |
| 6 | Sonnet 5.5 | WebMCP | 90.0 | 49/49 | 147/147 (100.0%) | $0.003530 | 5,046 | 4.2s |
| 7 | GPT-6.1 Sol | WebMCP | 85.4 | 49/49 | 147/147 (100.0%) | $0.003432 | 2,502 | 8.0s |
| 8 | Gemini 3.8 Flash | WebMCP | 85.0 | 49/49 | 147/147 (100.0%) | $0.003638 | 3,799 | 8.3s |
| 9 | GPT-6 Astra | WebMCP | 82.2 | 49/49 | 146/147 (99.3%) | $0.017100 | 2,575 | 6.3s |
| 10 | Muse Wonder | WebMCP | 80.6 | 49/49 | 146/147 (99.3%) | $0.015549 | 5,236 | 8.2s |
| 11 | Claude Opus 5 | WebMCP | 80.0 | 49/49 | 147/147 (100.0%) | $0.014013 | 4,770 | 9.8s |
| 12 | GPT-6.1 Sol | Code execution | 75.4 | 49/49 | 147/147 (100.0%) | $0.019001 | 10,503 | 16.5s |
| 13 | GPT-6 Luna | Computer use | 72.6 | 41/49 | 128/147 (87.1%) | $0.002112 | 21,177 | 19.4s |
| 14 | Haiku 5.5 | Computer use | 72.5 | 41/49 | 123/147 (83.7%) | $0.001698 | 37,601 | 16.0s |
| 15 | GPT-6 Astra | Code execution | 70.3 | 49/49 | 147/147 (100.0%) | $0.118676 | 10,982 | 16.4s |
| 16 | Sonnet 5.5 | Computer use | 69.8 | 47/49 | 140/147 (95.2%) | $0.031641 | 37,339 | 19.7s |
| 17 | GPT-6.1 Sol | Computer use | 66.3 | 45/49 | 134/147 (91.2%) | $0.039839 | 20,797 | 20.8s |
| 18 | Jev + Mercury 2.5 | DOM (ultrafast) | 63.4 | 25/49 | 76/147 (51.7%) | $0.000778 | 13,892 | 5.4s |
| 19 | GPT-6 Astra | Computer use | 61.5 | 45/49 | 135/147 (91.8%) | $0.260780 | 20,560 | 20.8s |
| 20 | Claude Opus 5 | Computer use | 56.4 | 45/49 | 134/147 (91.2%) | $0.138857 | 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.
We operate eight real websites across a wide range of tasks — 49 tasks, three attempts each. There are four ways to operate the website:
The website exposes direct actions for the agent to call.
The agent reads rendered images of the page and acts by coordinate.
The agent reads the page's DOM and accessibility tree.
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.
complete_checkout WebMCP tool, fixing the checkout task on the online store.