Anthropic's Fable 5.2 completes 41 percent of real freelance jobs, up from 2.5 percent a year ago

Anthropic's new AI model, Fable 5.2, has made a huge improvement on the Remote Labor Index. It completes more than 41 percent of real freelance projects at a professional level. This puts it far ahead of its previous version and all other competing models.

On the Remote Labor Index (RLI), created by Scale Labs and the Center for AI Safety (CAIS), Anthropic's new model Fable 5.2 ranks first with a score of 41.3 percent. The test measures how often AI agents can complete real freelance jobs well enough for a paying customer to accept the work.

This shows how quickly AI agents are improving at practical work. When the test started about a year ago, the best system, Manus, completed only 2.5 percent of the tasks. Fable 5.2 now completes more than 16 times as many jobs at a professional level.

The improvement is also large compared with recent results. In July, Fable 5, an older model from the same family, was still in first place with about 16 percent. That was already a major step forward: the best AI agents had more than quadrupled the share of jobs they could complete at a professional level in eight months. Just four months later, Fable 5.2 reached 41.3 percent. That is about two and a half times Fable 5's July score. Its direct predecessor, Fable 5.1, scores just under 18 percent. This means Fable 5.2 completes more than twice as many jobs successfully.

The next two places on the current leaderboard belong to Anthropic's Opus 5.5, with 21.25 percent, and OpenAI's GPT 6 Astra, with 20.83 percent. Fable 5.2 is almost twice as successful as the next-best models. Google's strongest model in the test, Gemini Argon 4, comes fifth with 17.08 percent. Older models, such as GPT-5.2 and Gemini 3 Pro, score less than three percent.

Fable 5.2 makes big improvements in 3D models, audio, and architecture

Scale Labs said on X that Fable 5.2 has improved most at editing audio and music, designing buildings and indoor spaces, and creating 3D models. The AI agent can also handle long and difficult projects that top-ranked models in earlier tests often could not finish.

Almost 50 jobs that Fable 5.2 completed successfully had never been completed by any other AI agent in the test. However, Scale Labs says the model still struggles with other kinds of work, especially designing industrial products and creating animations. According to the researchers, when AI agents fail, they usually produce work that looks amateur, unfinished results, files that do not work, or designs that contradict themselves.

Real jobs worth $144,000

Unlike many standard AI tests, which check only one skill at a time, the RLI research paper looks at complete work projects. Its dataset contains 240 finished projects from 358 verified professionals on the Upwork platform. These projects are worth about $144,000 in total.

Each project costs an average of $632.60 and takes a human freelancer just under 29 hours to complete. The jobs cover many kinds of work, including video editing, computer-aided design (CAD), game development, and data analysis. AI agents work on these projects inside a virtual computer workspace. It includes common professional programs such as Blender, GIMP, and Audacity, as well as special AI tools for creating images, audio, and video.

Trained human experts manually check whether an AI agent has completed a project successfully. They open the submitted files in the correct professional software and compare the AI's work directly with the work done by the human freelancer. CAIS has tried to use another AI system to do this time-consuming checking, but these attempts have not worked well so far. AI reviewers have trouble using professional software themselves. As a result, they can be fooled by work that looks good at first but has technical problems. This leads them to give AI models much higher scores than they deserve.