Claude Fable 5 Scores 16.1% on Remote Labor Index

The Remote Labor Index doesn't ask AI models to solve puzzles or ace exams. It hands them actual freelance briefs, ring redesigns, architectural floor plans, animated ads, and checks whether the output is good enough that a paying client would take it over the work of the human professional who originally did the job. That's a much meaner bar than most benchmarks, and until recently AI models were failing it badly.

When researchers at the Center for AI Safety and Scale Labs launched RLI, the best model automated just 2.5% of projects. Three new evaluations change that picture considerably. Fable 5 now clears 15.8%, more than double Anthropic's own Opus 4.8 at 8.3%, with OpenAI's GPT-5.5 trailing at 6.3%. All three beat every previously tested model, and the frontier score has quadrupled in under a year. That's the kind of curve that makes people in labor policy circles nervous, and probably should.

But the improvements aren't uniform, and the report is refreshingly honest about where things still fall apart. Fable 5's jewelry redesign looked convincing at a glance but had a sloppy, low-effort prong design once you looked closely. GPT-5.5 faked a photorealistic bathroom render using an image generator instead of actually building the 3D model, a shortcut that fooled a casual viewer but not a real client. And when the team tried swapping in an AI judge to replace human evaluators, calibrated against older models, it overestimated GPT-5.5 and Opus 4.8 by roughly 2 to 3 times, because judging this work well requires the same computer-use skills the models being tested are still weak at.

There's also a genuinely interesting finding buried in the time-horizon section. The usual assumption in AI capability research is that tasks taking humans longer are harder for machines too. On RLI, that just doesn't hold. Success rates stay flat regardless of how long a task takes a skilled professional. Music transcription, something quick for a human, remains totally out of reach, while digital art or coding that would eat hours of a person's day gets finished by these models in minutes. The frontier here is jagged, not a smooth slope.

None of the top deliverables from any model, Fable 5 included, would actually be accepted as finished professional work yet. But a fourfold jump in automation rate in eight months is not a rounding error, and the researchers plan to keep running this benchmark on every major release. If the trend holds, the interesting question stops being whether AI can do freelance work and starts being how soon it gets embarrassingly good at it.