Skip to content
Sıfırıncı Dakika
Latest

Claude Fable 5.2 builds a working operating system with 1,000 agents in eight days

ModelsNew2 min read

In brief

Anthropic shared internal test results showing the capabilities of Claude Fable 5.2 and Mythos 5.2 across software, robotics and cybersecurity. Fable 5.2 coordinated more than 1,000 subagents over eight days to build an operating system that boots on a laptop. The company says the models generate output up to 30 percent faster than their predecessors.

  • Built a working OS by coordinating over 1,000 subagents
  • 64.9 percent success rate on RoboDojo
  • 89.6 percent score on CVE Rediscovery with Mythos 5.2
  • Up to 30 percent faster output than predecessors
  • Pricing from $8 per million input tokens and $40 per million output tokens
  • Optional fast mode with up to 2.5x speed increase

An operating system built by agents

Anthropic's new models, Claude Fable 5.2 and Claude Mythos 5.2, took on complex engineering tasks in the company's internal tests. In the most striking experiment, an engineer gave Fable 5.2 a basic prompt and an empty project folder. Over eight days, the model coordinated more than 1,000 subagents that wrote code, tested components and reviewed each other's work.

The resulting operating system could boot on a laptop. It included its own kernel, hardware drivers and permanent file storage, along with a graphical interface, login screen, installer and built-in Claude functionality.

Hardware and software tasks

Fable 5.2 designed three electronic devices: a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. The model selected components from US distributors, designed circuit boards, created 3D-printed enclosures and wrote the required software. All three devices worked without electrical modifications.

In a separate internal migration, a team of agents moved a critical production service from Python to Rust. The resulting service now starts roughly 26 times faster.

Robotics and cybersecurity

On the RoboDojo benchmark, Fable 5.2 reached a 64.9 percent success rate, well above Fable 5.1's 27.6 percent. GPT-6 Astra scored 29.8 percent on the same test, while Opus 5.5 reached 39.4 percent. The benchmark measures how models control a simulated two-arm robot using camera images.

The security-focused Mythos 5.2 scored 89.6 percent on the CVE Rediscovery evaluation. That is a 10.2-point improvement over Opus 5.5 and an 11.3-point increase over Mythos 5.1's 78.3 percent.

Pricing and speed

According to Anthropic, both models generate output up to 30 percent faster than their predecessors. Standard pricing starts at $8 per million input tokens and $40 per million output tokens. An optional fast mode increases token generation speeds by up to 2.5 times but doubles the standard price. The company also says its safety classifiers have improved, with Fable 5.2 producing about 55 percent fewer unnecessary blocks than Fable 5.

Why it matters

This report shows with concrete experiments that frontier models can take on multi-step engineering work, not just chat. It helps readers understand the shifting limits of AI in software and hardware production.

Sources

Related stories