Anthropic's Fable 5.2 Completes 41% of Real Freelance Jobs, Up From 2.5% a Year Ago

Anthropic's newly released model Fable 5.2 scored 41.3% on the Remote Labor Index (RLI), a benchmark that asks AI agents to complete real freelance jobs on Upwork and grades their work against human deliverables. The result, released by Scale Labs and the Center for AI Safety (CAIS), means a paying client would accept the AI's work on roughly 4 out of every 10 professional projects, according to The Decoder.

The score more than doubles the next‑best system. Anthropic's own Opus 5.5 scored 21.25%, and OpenAI's GPT 6 Astra came in at 20.83%. Google's Gemini Argon 4 managed 17.08%, according to the Scale Labs RLI leaderboard.

When the RLI benchmark was first published in October 2025, the best system — a general‑purpose agent called Manus — completed just 2.5% of projects. By July 2026, Anthropic's Fable 5 had reached about 16%. Now, five months later, Fable 5.2 has pushed that to 41.3% — a 16x improvement in roughly 13 months, as tracked by MLQ.ai.

"Fable 5.2 now completes more than 16 times as many professional jobs," The Decoder reported. The acceleration curve is steep: the jump from Fable 5.1 (under 18%) to Fable 5.2 (41.3%) happened in a single generation.

The RLI uses 240 completed Upwork projects from 358 verified professionals, worth approximately $144,000 total. The average project cost $632.60 and took a human about 29 hours to complete,.¹ Tasks span video editing, CAD design, game development, data analysis, and more.

AI agents work in a virtual environment with professional software like Blender, GIMP, and Audacity. Crucially, human experts grade every submission by opening the files in the relevant professional software and comparing them directly to the human freelancer's work. ¹ that AI reviewers cannot operate the professional software themselves and are easily fooled by superficially polished but technically flawed outputs — making human grading essential.

Scale Labs told The Decoder that the largest gains were in audio and music editing, architecture and interior design, and 3D modeling. Nearly 50 of Fable 5.2's successful completions were tasks no AI agent had ever solved before in this benchmark.

The 41.3% figure is both impressive and alarming, depending on your vantage point. On one hand, it means AI agents can now handle nearly half of a representative sample of freelance work at professional quality — substantially more than most people would have predicted a year ago. On the other hand, 58.7% of jobs remain out of reach, with common failure modes including amateur‑quality work, incomplete results, broken files, and contradictory designs, as.¹

For builders and developers, the implication is clear: AI agents are becoming viable for production freelance work in specific domains, particularly structured creative tasks like 3D modeling and audio editing. The trajectory — from 2.5% to 41% in 13 months — suggests the remaining 59% gap may close faster than the first 40% did.

Fable 5.2's 41.3% nearly doubles Opus 5.5 (21.25%) and GPT 6 Astra (20.83%), giving Anthropic a commanding lead on the RLI, according to the Scale Labs leaderboard. This is notable because Fable 5.2 is Anthropic's frontier‑tier model priced at $8 per million input tokens and $40 per million output tokens — double the cost of Opus 5.5 ($4/$20 per million tokens), per Anthropic's official pricing. The RLI result was included in Anthropic's November 9 launch materials for Fable 5.2 and Mythos 5.2, alongside other benchmark scores, though the independent Scale Labs leaderboard provides more detailed breakdowns.

The pricing dynamic is striking: Anthropic's most expensive model outperforms its own Opus line by nearly 2x on real‑world task completion, but it costs twice as much to run. For builders deciding which model to deploy for agentic workflows, the RLI suggests the premium for Fable 5.2 translates into meaningfully better results on complex, multi‑step freelance‑style projects. The fact that the next two slots are held by Anthropic's Opus 5.5 and OpenAI's GPT 6 Astra — both within a point of each other at ~21% — shows that while the frontier is moving fast, the gap between first and second place remains substantial.