China’s Monica startup launched Manus AI on March 6, 2025, introducing a fully autonomous agent that plans and executes multi-step tasks independently. The system utilizes a multi-agent architecture within a Linux sandbox to perform web automation, code execution, and data processing.
- Manus achieved scores of 86.5%, 70.1%, and 57.7% on GAIA benchmark levels 1, 2, and 3 respectively.
- These results surpass OpenAI’s Deep Research system, which scored 74.3%, 69.1%, and 47.6% across the same levels.
- The model outperformed the previous state-of-the-art baseline of 67.9% on Level 1 tasks.
- Access is currently limited to an invitation-only beta phase via the official website.
The benchmark performance suggests Manus may be one of the most capable autonomous agents available, though real-world usability depends on handling unpredictable tasks.