Claude Fable 5.2 Sets Record on Freelance-Task Benchmark

The Remote Labor Index, a test built by Scale Labs and the Center for AI Safety (CAIS), measures how many real freelance jobs from the freelance site Upwork an AI model finishes well enough for a paying client to accept the work. Anthropic's new model, Fable 5.2, completed 41.3 percent of these jobs, almost twice the score of Anthropic's own Opus 5.5 (21.25 percent) and OpenAI's GPT-6 Astra (20.83 percent). A year ago, the top performer on the test, an AI agent called Manus, managed only 2.5 percent.

That 16-fold jump in a year shows how quickly AI agents are learning ordinary freelance office work. But the same test shows a clear limit. Fable 5.2 still struggles with industrial product design and animation, often turning in unfinished or contradictory files. That suggests companies are better off giving such agents narrow tasks for now, rather than handing them an entire freelance project.