Anthropic's Fable 5.2 scores 41.3% on Remote Labor Index, up from 2.5% top score a year ago

The Remote Labor Index leaderboard, run by Scale Labs and the Center for AI Safety, now lists Anthropic's Fable 5.2 at 41.3%. That is the share of 240 real Upwork-sourced freelance projects (worth about $144,000 in total) completed at a level a client would accept, as judged manually by human experts. Earlier scores on the same benchmark were Fable 5.1 just under 18%, Opus 5.5 at 21.25%, GPT 6 Astra at 20.83%, Gemini Argon 4 at 17.08%, and Manus at 2.5% when the benchmark launched about a year ago. The reported gains are largest in audio and music editing, architecture and 3D modelling.

The benchmark is run by third parties and uses paid, human-graded work rather than synthetic tests, so the roughly twofold lead over the next-best models is more informative than a vendor's own claim. Freelancers in the named categories, and buyers of that work, now have a concrete reason to test current agents on their own routine projects. Still, 240 tasks is a small sample, and "acceptable quality" is a judgement call, so the result shows direction rather than a measured effect on jobs or pay. It also says nothing about cost, speed, or how much human checking is needed in practice.

Phrases like "massive leap" and the "16x" comparison lean on a year-old 2.5% baseline from a different kind of system, and the same article's table shows rivals clustered around 17 to 21%. "Completes 41 percent of real freelance jobs" overstates it: these are 240 curated test projects, not the live market. The coverage also does not link the leaderboard, paper or Scale Labs' post, so the numbers are not directly checkable from it.

1. 01Whether the leaderboard and paper confirm 41.3% on the primary source, including the grading method, any project-level breakdown, and whether the evaluation was run independently or on Anthropic-supplied outputs.
2. 02Whether OpenAI and Google's next releases close the gap on the Remote Labor Index, and whether other independent evaluators reproduce a similar jump on different real-work tasks.
3. 03Evidence outside benchmarks: changes in Upwork posting volumes, rates or completion patterns in audio editing, architecture and 3D modelling over the next two to three quarters, plus cost per completed project versus a human freelancer.