Julian Horsey writes at Geeky Gadgets about Fable 5.2, an AI model from Anthropic launched alongside Mythos 5.2, and a test where it built an operating system from scratch. Starting from an empty folder, the model coordinated more than 1,000 agents running for eight days. The result was a working prototype with drivers, file storage and a Windows-style installer. The report draws on coverage by Wes Roth.
In testing, Fable 5.2 solved multi-step problems that other models, including Opus 5.5, had trouble with, and it’s described as strong at robotic control and hardware engineering. Cost per task is lower than Opus 5.5 in some cases. In another test, it improved a game’s graphics in about an hour, adding 2D sprites, camera zooms, replays and scores, with one instance working on the engine and another on the graphics.
Anthropic added limits during training to hold back recursive self-improvement, where a model improves itself.…