Anthropic's Fable 5.2 completes about 41% of a set of 240 real freelance projects at a quality a paying client would accept. A year ago the best agent completed 2.5%, and in July the best was at about 16%. The people this reaches are freelancers, and the businesses that hire them, in fields like audio, architecture and 3D.
The jobs that go first are the ones you can check
Every project in the test has a written brief, the input files, and the finished work a human freelancer delivered and the client accepted. To grade an agent, trained people open its files in the same professional software, such as Blender or Audacity, and compare them with the freelancer's version. Researchers tried to have another AI do that grading and it did not work: AI reviewers cannot operate the software well, so they are fooled by work that looks polished and is technically wrong.
That is the useful lesson for anyone outside the lab. AI takes over work first where the result can be checked against a clear standard, and the checking is still done by a person. It is the same question we asked in which work AI takes over first, answered here with ordinary jobs instead of mathematics.
What a buyer can do with it now
Take a small online shop that needs a 40-minute supplier interview cleaned up for its podcast, and a floor plan for a pop-up showroom. Until now it posted each job, waited a few days and paid a freelancer. The median job in the test was worth $200 and took a freelancer about 11 hours. Now the owner can give each one to an agent first, open the result, and post the job only if it fails.
Say four in ten come back acceptable. Of ten jobs worth $200 each, that is about $800 not spent, and the six that fail cost perhaps 15 minutes each to open and reject. For audio, architecture and 3D the odds are probably better than four in ten, and for animation they are worse. Upwork has already built a way for agents in Claude, ChatGPT and Cursor to hire human freelancers, so the work that fails can flow back to people without the owner doing anything.
The first wave looked small, and this one reaches more fields
After ChatGPT came out in November 2022, writing freelancers on Upwork saw monthly jobs fall 2% and monthly earnings fall 5.2% in the following months. That looked mild, and the more experienced writers lost more than the beginners. In 2026 the marketplace itself cut its forecast, and it says about 10% of its volume is exposed to AI.
An agent that went from 2.5% to 41% of finished jobs in a year, across fields that were barely touched before, will cause bigger moves than that first wave did. The freelancers most at risk are those whose work is to take a brief and return a file. The safest are those who talk to the client, decide what the brief should have said, and check the result, because that is the part the test still needs a person for. If your own work is a brief in and a file out, the question to ask is who checks the file, and whether that could be you.