By Hamza Faisal

Anthropic's newest Claude model has pulled off something no AI system has done quite like this before: build a working operating system almost entirely on its own. In an internal experiment, Claude Fable 5.2 coordinated more than 1,000 AI subagents over eight days to design, write, test and review the code for a prototype operating system — starting from nothing but an empty folder and a single instruction.

The result boots on a real laptop. According to Anthropic, the system includes its own kernel, drivers for disk, keyboard and network hardware, persistent file storage, windows, a login screen and an installer, along with built-in Claude functionality. It is a prototype, not a rival to Windows, Linux or macOS, but the experiment is one of the clearest demonstrations yet of AI agents working together on a long, complex engineering task that would normally take a human team months.

A thousand agents, one project

The experiment began when an Anthropic engineer asked Fable 5.2 to build a prototype operating system from scratch. Rather than producing one long response, the model organised a team of more than 1,000 subagents, each handling a different part of the project. Over eight days, those agents wrote code, tested components and reviewed each other's work, with minimal human intervention along the way.

The same approach scaled to other demanding problems. A team of 1,000 Fable 5.2 agents formalised the proof of Fermat's Last Theorem into a computer-checkable format in 30 hours — a task that took an earlier research version of Claude 11 days. The model also rewrote a software service from Python into Rust, with Anthropic saying the rewritten service ran around 26 times faster, and designed three working electronic devices: a home voice-assistant hub, a USB speakerphone and a handheld virtual pet.

What the new models can — and cannot — do

The experiments arrived with the launch of Claude Fable 5.2 and Claude Mythos 5.2 on November 9. Both share the same underlying model but carry different safeguards: Fable 5.2 is broadly available, while Mythos 5.2 is restricted to vetted organisations working in cybersecurity and life sciences. Fable 5.2 scored 49.1 percent on the FrontierCode 2.0 benchmark for agentic software development, up from 32.4 percent for Fable 5.1, and Anthropic says it generates responses up to 30 percent faster. API prices were cut by 20 percent, to $8 per million input tokens and $40 per million output tokens.

There are important caveats. These are company-reported results that have not been independently verified, and the prototype operating system has not been shown to meet the security, compatibility, reliability or maintenance standards of a mature system. Anthropic itself notes that reliability remains an issue without human supervision, so the demonstration is a sign of how far agent teams have come rather than proof they can work unsupervised.

Why it matters

The shift on display here is from AI that answers questions to AI that sustains real work. A single model coordinating a thousand agents across eight days points toward software teams, hardware labs and research groups handing whole projects — not just individual tasks — to agent systems. For enterprises, that could mean faster development cycles and cheaper experimentation, but only if the reliability gaps close. For now, the operating system that a thousand agents built in eight days is best read as a preview: the tools are getting capable enough to build serious things, and the supervision problem is now the bottleneck.