Anthropic’s Claude Fable 5.2 Finishes 41% of Real Freelance Jobs, Beating Every Other AI Model
Anthropic's Claude Fable 5.2 just topped a test that measures how well AI can do real paid freelance work, finishing 41.3% of 240 actual jobs well enough to pass as human-quality work.
What this means for you
If you hire freelancers, use AI tools, or just want to know how far this technology has come, this result matters. It comes from a benchmark called the Remote Labor Index, or RLI. Think of RLI as a report card that uses real freelance jobs instead of made-up tests.
Here is what happened. Claude Fable 5.2 scored 41.3% on RLI. That means out of 240 real freelance projects pulled from Upwork, human experts judged Fable 5.2's work to be as good as or better than the original human freelancer's work on 41.3% of them. No other AI model came close. Anthropic's own Opus 5.5 placed second at 21.25%, and OpenAI's GPT 6 Astra came in third at 20.83%.
This score was not part of Anthropic's official announcement when it launched Fable 5.2 and another model called Mythos 5.2 on November 9. Instead, the number comes from the people who run the benchmark: Scale Labs and the Center for AI Safety, known as CAIS.
How fast scores have climbed
A year ago, this kind of score would have sounded impossible. When the RLI benchmark first launched in October 2025, the best AI system at the time, called Manus, could only finish 2.5% of the jobs. By July, Anthropic's Fable 5 had pushed that number up to about 16%. Its successor, Fable 5.1, reached 17.92%. Now Fable 5.2 has more than doubled the second-place score.
Here is the full leaderboard, according to Scale Labs
- Fable 5.2 (Anthropic): 41.30%
- Opus 5.5 (Anthropic): 21.25%
- GPT 6 Astra (OpenAI): 20.83%
- Fable 5.1 (Anthropic): 17.92%
- Gemini Argon 4 (Google): 17.08%
- Fable 5 (Anthropic): 15.83%
Reporting from The Decoder found that Fable 5.2 made its biggest gains in audio and music editing, architecture and interior design, and 3D modeling. Nearly 50 of the jobs it completed had never been finished successfully by any AI model before. Still, the model struggles with industrial product design and animation. Scale Labs says the AI "can barely handle" those areas, often turning in work that looks amateur, is missing pieces, or has designs that contradict each other.
How the test actually works
The 240 jobs in this benchmark come from 358 real, verified Upwork freelancers. Together, the jobs are worth about $144,000, averaging $632.60 and about 29 hours of work each. The work covers things like video editing, CAD design, game development, and data analysis.
When an AI model takes the test, it works inside a virtual desktop using real professional tools like Blender, GIMP, and Audacity, the same tools a human freelancer would use. The leaderboard rankings are based on 230 of these projects kept private so AI makers cannot train their models to game the test.
Grading is done by trained human experts, not by another AI. They open each AI-made file in the right software and compare it directly to what the original human freelancer delivered. CAIS has tried to get AI to do this grading instead, but so far that has not worked well. The AI reviewers have trouble operating the professional software themselves, and tend to grade too generously.
It is worth being clear about what this score does and does not tell you. It measures how well AI did on this one set of 240 real jobs. It does not tell you what percentage of all freelance work worldwide AI could take over. Anthropic has not publicly commented on the result.
If you want to understand how close AI really is to doing skilled human jobs, don't just watch the splashy product launches. Keep an eye on benchmarks like this one, which test AI against real paid work and real human judges, not just marketing claims.
Source: MLQ.ai
