New research from the Center for AI Safety and Scale AI found that leading AI agents perform poorly when tested on real-world freelance tasks. Across multiple industries, including game development and data analysis, the models completed less than 3% of assigned work, generating only $1,810 out of a potential $143,991 in value.

Even top-performing systems like Manus, Grok 4, and Claude Sonnet 4.5 achieved automation rates between 2.1% and 2.5%. The findings highlight a major gap between AI hype and actual productivity, especially when compared to human freelancers handling complex, real-world tasks.