One freelance job in six

The Remote Labor Index is a benchmark built from 240 freelance projects that were genuinely commissioned and paid for on the open market — 3D modelling, video edits, graphic design, small web apps — a median of eleven and a half hours' work each, roughly $200 a job, about $144,000 of real human labour in total. The test is deliberately unforgiving: a deliverable counts as a success only when two of three human experts, reviewing it holistically, would sign off on it. No unit tests, no synthetic tasks, no partial credit for looking plausible.

When the index launched in October 2025, the best agent cleared 2.5% of it — six jobs out of 240, worth about $1,720 of the pile. The first frontier refresh since then puts Claude Fable 5 at 16.1%. Opus 4.8 sits at 8.3%, GPT-5.5 at 6.3%. Because the task set is fixed and priced to real market rates, the jump is a like-for-like reading rather than a moved goalpost: on the same jobs, agents finish more than six times as much acceptable paid work as they did eight months ago.

One detail makes the measurement oddly concrete. Twenty-two of the 240 projects couldn't be run on Fable 5 at all — the US export-control order that Anthropic answered by shutting the model down worldwide left a literal hole in the benchmark. The leaderboard notes that even if Fable 5 had failed every one of those missing jobs, it would still score 14.6%, still comfortably ahead. The policy story Mentat covered as a shutdown shows up here as absence.

The number to hold onto is the one that isn't in the headline: five jobs in six still come back rejected. This is a trajectory, not a takeover — a curve steep enough that the people who commission this work, and the people who price it, now have a real figure to watch instead of a vibe.