Anthropic's Fable 5.2 completed 41.3% of real freelance projects at professional quality on the Remote Labor Index, nearly doubling the rates posted by Claude Opus 5.5 at 21.3% and GPT-6 Astra at 20.8%, according to results reported by The Decoder1.
The Remote Labor Index, developed by Scale Labs and the Center for AI Safety (CAIS), measures how often AI agents can finish end-to-end freelance jobs rather than isolated skill tests2. The benchmark draws on 240 self-contained projects sourced from 358 verified freelancers on the Upwork platform, spanning 23 of Upwork's 64 domains. Each project includes the original brief, input files, a professionally accepted human deliverable, and economic data on human completion time and cost. Projects requiring physical labor, long-term evaluation, or direct client interaction are excluded.
Fable 5.2 posted particular strength in 3D models, audio, and architecture categories. The gap over Opus 5.5 and GPT-6 Astra is wide: Fable 5.2's 41.3% completion rate is roughly double either rival's score.
ANALYSIS The margin is notable given that Opus 5.5 is Anthropic's own flagship reasoning model, meaning the newer, lighter-weight Fable line outperformed it on practical work tasks by a factor of nearly two.
Anthropic released Fable 5.2 on November 9, positioning it for what the lab called open-ended technical work[1]. The RLI results arrive one day after that launch, offering an independent, economically grounded check on the model's capabilities[1].
Separately, GPT-6 Astra recently topped Alibaba's Co-Design Bench, a coding-agent benchmark built from real user requests, but fully completed only 26.4% of 273 tasks there[2].
ANALYSIS Across both benchmarks, the pattern is consistent: frontier models still fail to finish the majority of real-world professional tasks, even as Fable 5.2's 41.3% marks a step-change from the roughly one-in-five completion rates of its closest competitors.
Scale Labs describes the RLI as addressing a gap left by research benchmarks that test isolated skills, providing instead a direct comparison of AI agent output against professional human deliverables.