mlq.ai reports that Claude Fable 5.2 scored a 41.3% automation rate on the Remote Labor Index (RLI), a benchmark where AI agents attempt actual freelance jobs, with a score meaning expert human graders judged the deliverable at least as good as the human freelancer's work.

Anthropic's Opus 5.5 is second at 21.25% and OpenAI's GPT 6 Astra third at 20.83%, meaning the new model roughly doubled the previous best.

The score comes from the benchmark's operators, Scale Labs and the Center for AI Safety (CAIS), and was not part of Anthropic's launch materials for the model, which also debuted at No. 1 on Artificial Analysis' Intelligence Index, edging Google's unreleased Gemini 4 Carbon by 5 points.

The pace of improvement is the story. When the RLI paper was published in October 2025, the best system, Manus, completed just 2.5% of projects.

Fable 5 led at about 16% in July, and its direct predecessor Fable 5.1 scored 17.92%. Moving from under 3% to over 41% in roughly a year — and from 17.9% to 41.3% in a single model update — is a leap of the kind benchmarks on real economic tasks have rarely shown.

According to mlq.ai, nearly 50 of Fable 5.2's successful jobs were ones no agent had ever completed before, and Scale Labs saw the largest gains in audio and music editing, architecture and interior design, and 3D modeling.

Unlike benchmarks graded by AI judges, RLI's 240 projects came from 358 verified Upwork freelancers and are worth about $144,000 in total — an average of $632.60 and roughly 29 hours of human work each, spanning video editing, CAD, game development and data analysis.

Agents work in a virtual desktop with tools such as Blender, GIMP and Audacity.

mlq.ai reports that trained experts open each submission in the relevant professional software and compare it against the freelancer's delivered work; CAIS's attempts to automate grading with AI reviewers have so far failed because the reviewers struggle to operate the software and grade too generously.

The leaderboard itself uses 230 private projects.

The record is not a clean sweep: mlq.ai, citing The Decoder, reports that industrial product design and animation are areas the AI 'can barely handle,' with typical failures including amateur-quality work, incomplete or broken files, and contradictory designs.

The benchmark's operators also caution that scores measure completion on a fixed sample and do not estimate what share of the remote labor market AI can automate, and Anthropic has not commented on the result.

Still, the implication for freelance platforms and clients is hard to ignore — an agent that a reasonable client would accept on over 4 in 10 projects changes the economics of exactly the task types RLI tests.

It also lands amid Anthropic's own 'pacing, not pausing' positioning, with 5.2 priced 20% below 5.1 at $8 in / $40 out per million tokens.

Anthropic's 'pacing' strategy is now visible in practice: executive Dianne Penn said of the company's deliberate slowdowns, 'We called for pacing, not pausing,' yet the still-regulated Fable sits atop both a real-work benchmark and the Artificial Analysis leaderboard.

With OpenAI expected to release a new model and Google's Gemini 4 Carbon still unreleased, the next rounds of RLI will show whether the 41.3% figure is a plateau or a floor.

Builders watching the agent space should note that agentic coding remains tiered — Anthropic's own Mythos 5.2 is restricted to vetted security and life sciences groups — and that Fable represents the guardrailed public frontier of the same model family.