AI agents now finish 16 percent of real freelance jobs, up from 2.5

The Center for AI Safety has updated its Remote Labor Index, and the numbers moved fast. The benchmark takes real freelance jobs that human professionals were actually paid for, across areas like 3D modeling, architecture, graphic design, video, audio, data analysis, and web development, then checks whether an AI agent can deliver work a human reviewer would accept. When the index launched, the best agent finished 2.5 percent of the projects. The latest run puts Fable 5 at 15.8 percent, about double Opus 4.8 at 8.3 percent and ahead of GPT-5.5 at 6.3 percent. The top score has more than quadrupled in under eight months.

One detail is worth reading closely. The researchers also tried scoring the new models with an automated LLM judge that had been calibrated on older systems, and it inflated the results badly: roughly 2.9 times too high for GPT-5.5 and 2.3 times too high for Opus 4.8. It still ranked the models in the right order, but the absolute numbers were far off, and human evaluators caught the gap. Most freelance projects still defeat these agents. The direction and speed of the curve are the point.

Why it matters

If you hire or sell freelance digital work, a jump from 2.5 to nearly 16 percent in eight months is a trend line to plan around now. And if you evaluate agents with an LLM judge, recalibrate it against fresh human review, because a judge tuned on last year's models can overstate today's by more than double.