AI agents are making a larger mark on complex freelance work. On the Remote Labor Index, Anthropic’s Fable 5.2 completed 41.3 percent of projects at a quality level suitable for acceptance by a paying client, moving well ahead of the other systems measured.
A sharp rise on a real-work benchmark
The Remote Labor Index (RLI), from Scale Labs and the Center for AI Safety (CAIS), evaluates whether AI agents can complete entire freelance assignments, rather than demonstrate isolated skills. When the benchmark launched about a year ago, the leading system, Manus, completed 2.5 percent of tasks. Fable 5.2’s current result is more than 16 times that share.
The latest jump is substantial even against recent results. In July, Fable 5 scored about 16 percent. Four months later, Fable 5.2 reached 41.3 percent. Its direct predecessor, Fable 5.1, scored just under 18 percent, so the newer model completed more than twice as many jobs.
The gap extends to the rest of the current leaderboard. Anthropic’s Opus 5.5 scored 21.25 percent and OpenAI’s GPT 6 Astra scored 20.83 percent. Google’s highest-ranked model in the test, Gemini Argon 4, placed fifth at 17.08 percent. Older models such as GPT-5.2 and Gemini 3 Pro scored below three percent.
Where the model is gaining ground
Scale Labs said Fable 5.2 made its strongest gains in audio and music editing, architecture and interior design, and 3D modeling. It also completed long, complicated projects that top-ranked models had regularly failed in the past.
Nearly 50 of the model’s successful projects had not previously been completed by any agent in the test. That is a notable sign of progress on unfamiliar assignments, though it does not mean the system is broadly reliable across professional work.
Some categories remain difficult. Scale Labs said the system still struggles with industrial product design and animation. Across the benchmark, failed attempts commonly produce amateur-quality work, incomplete deliverables, broken files, or designs that contradict themselves. A polished-looking result can therefore fall short once someone opens and checks the actual files.
What the index asks AI agents to do
The RLI research paper describes a dataset of 240 completed projects from 358 verified professionals on Upwork. Together, the assignments are worth about $144,000. They cover varied work, including video editing, CAD design, game development, and data analysis.
A typical project costs an average of $632.60 and takes a human freelancer just under 29 hours. Agents work in a virtual environment with common professional software such as Blender, GIMP, and Audacity, as well as specialized AI tools for generating images, audio, and video.
This setup tests more than whether an AI can produce a plausible first draft. The agent must handle the assignment through to a usable deliverable in the relevant tools. The results suggest that current AI agents can complete a growing share of these end-to-end jobs, while also showing how much performance depends on the kind of work involved.
Why people still judge the results
Trained human experts review submissions by opening the files in the appropriate professional software and comparing the AI’s output with the freelancer’s completed work. That hands-on check matters because a result can appear convincing on the surface while containing technical problems.
CAIS has tried automating this review with another AI system, but those attempts have so far failed. The automated reviewers struggle to use professional software themselves, making them susceptible to polished but flawed submissions and leading them to rate models too generously.
For now, the benchmark’s human review helps distinguish a successful project from one that merely looks finished. Fable 5.2’s lead marks a clear improvement on the RLI, while the uneven results across categories and the difficulty of automated evaluation show why a high benchmark score is only part of the picture.