Claude Fable 5.2: 1,000+ AI agents build an operating system in eight days

What's new

Anthropic, the US company behind the Claude models, unveiled two new models on November 9, 2026: Claude Fable 5.2 and Claude Mythos 5.2. According to Anthropic, both share the same architecture. They differ in who can use them and how they're safeguarded. Fable 5.2 is aimed at a broad audience and is available through the API. Mythos 5.2 is reserved for vetted organizations working in cybersecurity or biomedical research.

Alongside the models, Anthropic showcased a series of internal experiments. The most striking: in eight days, Fable 5.2 built a prototype operating system almost from scratch, directing more than 1,000 software agents to do it. The results were reported by, among others, Russian site New-Science.ru and US magazine Interesting Engineering.

The operating system experiment

An Anthropic engineer gave Fable 5.2 a short brief and an empty project folder. The model then organized the work of more than 1,000 subagents. These are independently running AI instances that each take on a subtask, much like members of a large team. Around the clock, they wrote code, tested components, reviewed each other's work and fixed bugs.

After eight days, Anthropic says, the result was a system with these components:

• its own kernel, the core of the operating system that manages memory, processes and hardware
• drivers for disk, keyboard and network
• persistent file storage
• a window-based graphical interface
• a login screen and an installer
• built-in Claude features

The system reportedly booted on a real laptop and displayed a graphical interface. The limits matter, as New-Science.ru points out: Anthropic hasn't released the source code, independent test results or the cost of the experiment. So there's no way to judge how reliable and secure the prototype is, or how much different hardware it runs on. The operating system itself is less interesting, since it remains a demo project. What's interesting is that an AI system was apparently able to coordinate, over several days, the development of a large software project with many interdependent parts.

Anthropic cites one more software example: a team of agents ported an important internal production service from Python to Rust. Rust is a programming language known for being especially fast and memory-safe. According to Anthropic, the new service starts up about 26 times faster.

Hardware, math and physics

Together with the company Diode, Fable 5.2 designed three electronic devices: a voice assistant hub for the home, a USB speakerphone and a handheld virtual pet. The model selected components from US distributors and designed four-layer circuit boards, 3D-printable enclosures and the necessary software. According to Interesting Engineering, the boards were ready for manufacturing in less than a day. After fabrication, all three devices worked without any changes to the circuits. Human engineers did, however, assemble and test the devices.

In another experiment, a team of 1,000 Fable agents formalized the proof of Fermat's Last Theorem in Lean in 30 hours. Lean is a programming language for writing mathematical proofs in a way that lets a computer check every logical step. So the AI didn't find a new proof. It translated the known proof into a machine-checkable form, which until now has been considered extremely laborious work. The model also reportedly helped researchers with two previously unsolved problems in the theoretical physics of black holes. The sources give no details.

Benchmarks: robotics and security

Fable 5.2 has improved substantially at controlling robots. In the RoboDojo Sim benchmark, a model uses camera images to control a simulated two-armed robot. Demos included solving cube puzzles, unscrewing bottle caps and drawing pictures.

Mythos 5.2 scores 89.6 percent on the CVE Rediscovery test. CVEs are publicly cataloged security vulnerabilities in software. In this test, the model has to rediscover known vulnerabilities in old software versions on its own. Mythos 5.1 scored 78.3 percent, so the new version is 11.3 percentage points higher. Compared with Opus 5.5, Interesting Engineering cites a 10.2 percent improvement. An external security firm also found that other frontier models it tested missed one in six of the confirmed vulnerabilities Mythos 5.2 had found.

Speed, pricing and fewer refusals

According to Anthropic, both models generate responses up to 30 percent faster than their predecessors. API pricing:

Tokens are the chunks of text a model breaks language into, roughly a part of a word. The fast-mode prices are derived from the stated "double the standard price."

Anthropic has also reworked its safety filters, the classifiers: screening programs that sit in front of the model to detect and block risky requests. According to Anthropic, Fable 5.2 unnecessarily blocks about 55 percent fewer requests than Fable 5.1. For cybersecurity topics, refusals dropped by 60 percent, and for biology topics by 40 percent. If you do legitimate work in these fields, you should get blocked less often. Mythos 5.2 nonetheless remains available only to vetted organizations with approved projects.

Analysis

The results show where the big AI providers are heading: instead of a single chatbot, entire swarms of agents work on large projects for days. A bootable operating system in eight days, production-ready circuit boards in a day and a formalized Fermat proof in 30 hours would take human teams weeks or months. The jump on the robotics benchmark from 27.6 to 64.9 percent is also big, especially since Fable 5.2 is well ahead of GPT-6 Astra there.

That said, almost all the impressive numbers come from Anthropic's own experiments. It remains unclear how much compute time and money the 1,000 agents consumed, how many attempts were needed, how much humans intervened and how robust the result is. An operating system that boots once on a laptop is a long way from one that runs securely on varied hardware for years. Whether the technology holds up in practice can only be judged with open code, independent testing and experience from long-term use. For developers using the API, the near-term wins are mainly the higher speed and fewer false refusals. At $8 and $40 per million tokens, though, Fable 5.2 is among the pricier models.