Taskmosis Achieves #1 on Halluminate Web Bench
Leading AI web agents with 83.5% success rate — outperforming OpenAI Operator, Claude Computer Use, and all competitors while completing tasks in under 1 minute.
Top 3 Performers
Taskmosis Performance Breakdown
A New Benchmark Leader
We're excited to announce that Taskmosis has achieved the #1 position on the Halluminate Web Bench, the comprehensive benchmark for AI web agents.
With an overall success rate of 83.5%, Taskmosis outperforms every other AI web agent tested — including OpenAI's Operator (57.65%), Claude Computer Use (57.65%), and the previous leader rtrvr.ai (81.39%).
Speed Matters
Average task completion time in minutes. Taskmosis is up to 26x faster than alternatives.
Full Benchmark Results
| Agent | Overall | Read | Write | Time |
|---|---|---|---|---|
Taskmosis | 83.5% | 91.2% | 68.8% | 0.8 min |
rtrvr.ai | 81.39% | 88.24% | 65.63% | 0.9 min |
Browser Use | 59.77% | 67.65% | 43.75% | 4.3 min |
Claude Computer Use | 57.65% | 61.76% | 48.44% | 6.8 min |
OpenAI Operator | 57.65% | 61.76% | 48.44% | 20.7 min |
Convergence Proxy | 36.9% | 52.94% | 4.69% | 14 min |
Manus | 28.74% | 38.24% | 9.38% | 7 min |
Why Taskmosis Leads the Benchmark
Our architecture choices directly translate to benchmark performance. Here's why we outperform the competition:
CDP-Based Architecture
Our Chrome DevTools Protocol foundation gives us direct DOM access and native input events, eliminating the errors that plague screenshot-based agents.
Works in Your Browser
Unlike cloud-based agents, Taskmosis runs in your actual browser session with your cookies and credentials — no proxy authentication issues.
Text Model Efficiency
We process DOM structure with fast text models instead of expensive vision API calls. This makes us 10x faster and dramatically cheaper per task.
Precise Element Targeting
No more guessing pixel coordinates from screenshots. We identify exact elements semantically and interact with them directly via CDP.
Why Screenshot-Based Agents Struggle
Most competing agents rely on a screenshot → vision model → coordinate clicking pipeline. This approach has fundamental limitations:
- Vision model latency — 2-5 seconds per screenshot analysis
- Coordinate guessing — OCR and bounding box errors cause missed clicks
- Hidden element blindness — Cannot see elements outside the viewport
- High API costs — Vision calls are 10x more expensive than text
Taskmosis's CDP-based approach eliminates all of these issues by working directly with the page's DOM structure. Learn more about our architecture →
About Halluminate Web Bench
The Halluminate Web Bench is the industry's most comprehensive benchmark for evaluating AI web agents. It tests agents on real-world web tasks across diverse websites and scenarios.
READ Tasks
Information extraction tasks that require navigating to pages, finding specific data, and returning accurate results.
WRITE Tasks
Action-oriented tasks that require filling forms, clicking buttons, and modifying page state.
Frequently Asked Questions
Experience the #1 AI Web Agent
Try the benchmark-leading performance for yourself. 83.5% accuracy, 0.8 minute average speed, all running in your browser.