# Fable 5.2 completes 41% of real paid freelance jobs in a benchmark, up from 2.5% a year ago

> On the Remote Labor Index, built from paid Upwork projects worth about $144,000, Anthropic's new model roughly doubles the next-best agents. Humans still grade the work by hand.

- URL: https://devisdead.ai/news/fable-5-2-completes-41-of-real-paid-freelance-jobs-in-a-benchmark-up-from-2-5-a-
- Published: 2026-11-10T16:16:54Z
- Author: DevIsDead Newsroom
- Section: AI vs Humans
- Tags: anthropic, fable, benchmarks, freelance, agents
- Originally reported by The Decoder (https://the-decoder.com/anthropics-fable-5-2-completes-41-percent-of-real-freelance-jobs-up-from-2-5-percent-a-year-ago/).

Anthropic's Fable 5.2 now finishes 41.3% of the jobs in the Remote Labor Index (RLI), a benchmark made of real freelance work, Maximilian Schreiner reports for [The Decoder](https://the-decoder.com/anthropics-fable-5-2-completes-41-percent-of-real-freelance-jobs-up-from-2-5-percent-a-year-ago/). When the test launched about a year ago, the best agent, Manus, finished 2.5%.

## How fast that moved

- **A year ago:** Manus, 2.5%.
- **July:** Fable 5 led at about 16%.
- **Now:** Fable 5.2 at 41.3%, about two and a half times its July score and well over double Fable 5.1, which sits just under 18%.

The rest of the leaderboard trails far behind. Anthropic's Opus 5.5 is second at 21.25%, OpenAI's GPT 6 Astra third at 20.83%, and Google's best entry, Gemini Argon 4, fifth at 17.08%. Older models such as GPT-5.2 and Gemini 3 Pro stay under 3%.

## What the jobs are

RLI is built from 240 finished projects by 358 verified Upwork professionals, worth roughly $144,000 together. An average job paid $632.60 and took a human just under 29 hours: video editing, CAD, game development, data analysis. Agents work in a virtual workspace with tools like Blender, GIMP and Audacity, plus generators for images, audio and video.

According to Scale Labs, Fable 5.2 improved most at editing audio and music, architecture and interior design, and 3D modeling. Nearly 50 of its completed jobs had never been solved by any agent in the test. It still struggles with industrial product design and animation, and failed attempts usually look like amateur work, unfinished results, broken files or designs that contradict themselves.

## Humans still do the grading

Trained experts open each submission in the relevant software and compare it with what the human freelancer delivered. Efforts by CAIS, one of the groups behind the test, to hand that review to an AI have failed so far: AI graders can barely run the professional tools, so they're taken in by work that looks polished but is technically broken, and they score far too generously.

**[Scoreboard](/scoreboard) update:** AI +1. An agent now completes 41% of real paid freelance jobs, up from 2.5% a year ago.

Stay undead.