Claude Fable and Mythos 5.2 explained: Anthropic’s new model that wrote an entire operating system
Anthropic has rolled out Claude Fable 5.2 and Claude Mythos 5.2, which are two models designed to handle the toughest technical challenges. They are the same model but divided into two according to their safety measures; Fable is available to everyone while Mythos is accessible only to organizations involved in cybersecurity and life sciences. According to Anthropic, Fable 5.2 was able to coordinate over 1,000 subagents to create a bootable operating system from scratch and help formalize Fermat’s Last Theorem in just 30 hours. Here are the capabilities of these models, their cost, and availability.
One model, two products
Model-wise, both Fable and Mythos are the same. But it all lies in what they will be used for. Fable 5.2 has classifiers built in which restrict access to cyber activities which pose a risk such as penetration testing and exploit writing, as well as most biological research. Restricted requests will be directed to Opus models. According to Anthropic, the percentage of blocks with Fable 5.2 has been reduced by about 55 percent compared to Fable 5.1.
The operating system experiment
This main demonstration was carried out by a single Anthropic engineer who asked Fable 5.2 to develop a prototype operating system from scratch. The system coordinated 1,000+ sub-agents over a period of eight days to write, test, and review the code round-the-clock. This system can boot on a laptop and has a kernel, disk driver, keyboard driver, network driver, file system, windows, login screen, installation process and Claude capability. On another project front, a group of agents converted a live system from Python to Rust and its startup time improved by nearly 26 times.
Benchmarks and speed
As regards to FrontierCode 2.0, which examines how likely the code changes will be merged by open source maintainers, Fable 5.2 receives 49.1 percent on the Xhigh level of effort. According to Anthropic, Fable 5.2 ranks higher than Opus 5.5 (45.5%) and GPT-6 Astra (42.8%). These new models are also capable of producing outputs as much as 30% faster than the older 5.1 versions. The cost is $8 per million input tokens and $40 per million output tokens, with an optional fast version of the model at twice the speed for $16 and $80.
Where Mythos pulls ahead
Mythos 5.2 is the best Anthropic model for cybersecurity engineering. On the internal CVE Rediscovery benchmark, it achieved 89.6 percent and discovered 10.2 percent more vulnerabilities than Opus 5.5. According to a report by one security company, Mythos 5.2 detected bugs that each other frontier model could not detect in one out of six cases. Moreover, according to Mozilla, this model approximately doubled the vulnerability detection rate of Mythos 5.1 on their Firefox scans. In biology, Mythos 5.2 introduced an innovative way of predicting capsids of viruses, successfully predicting 28 out of 34.
Safety and what to use
Anthropic considers the probability of engaging in risky behavior in high-stakes environments to be low. For the sandbox escape experiment, Mythos 5.2 employed planted credentials in 3.1 percent of cases, compared to 21 percent with Mythos 5.1. Despite this, Anthropic recommends that its users begin with Opus 5.5 and progress to Fable or Mythos only for more complex assignments.
The Fable 5.2 model is now available on the Claude Platform under claude-fable-5-2 as well as Amazon Web Services, Google Cloud and Microsoft Azure. The enterprise version of Claude Security also uses Mythos 5.2 to conduct vulnerability scanning in code repositories. Access to Mythos 5.2 is limited to applying to the verification program.