One in six freelance jobs can now be completed by AI, start to finish, with no human involved. Eight months ago that number was one in 40.
Real Jobs, Real Money, No Toy Benchmarks
The Center for AI Safety runs something called the Remote Labor Index, and it skips synthetic tests entirely. They hand AI agents actual paid freelance work: 3D modeling, architecture, video, data analysis, the kind of jobs real clients pay real money for. Then they have working professionals grade the output against real deliverables, not a rubric written for a benchmark.
The Jump Is the Real Story
Claude Fable 5 completed about 16 percent of assigned projects start to finish. The best model eight months ago managed just 4 percent. That is more than a four times improvement in under a year, and the trend matters more than either number on its own.
Routine Work Goes First
The pattern in what AI wins and loses is consistent. It nailed accurate floor plans and clean 2D CAD drawings, the kind of work with a clear right answer. It still fumbles the craft: on a custom ring design, it got the overall shape right, but the prongs looked amateur next to a professional's work. Straightforward, clearly specified tasks fall first. Detail work that requires taste and judgment is still yours, at least for now.
AI Grading AI Lies
One more finding worth sitting with. When researchers tried using AI to grade AI generated work, it overstated the results by two to three times. Self grading is not a shortcut you can trust yet. Human review still matters, and skipping it means measuring a number that lies to you.
The Takeaway
The freelance work most exposed to AI right now is the routine, clearly specified kind. The work that survives is judgment heavy and hard to reduce to a rubric. Track which category your own work falls into, because the gap between 4 percent and 16 percent closed in eight months, and there is no reason to assume it slows down from here.