Anthropic Unveils Claude Fable 5.2 and Mythos 5.2 for Complex Technical Tasks

On November 9, 2026, Anthropic introduced updated artificial intelligence models Claude Fable 5.2 and Claude Mythos 5.2, designed for complex engineering, research, and software tasks. According to the developer, the new versions surpass their predecessors in several metrics of logical reasoning, code writing, computer vision, and scientific data handling. Both models use a common base architecture but differ in security mechanisms and access conditions.

Fable 5.2 is available to a wide range of users and has additional restrictions for potentially dangerous tasks in cybersecurity and biology. Mythos 5.2 is aimed at organizations engaged in professional software vulnerability searches and biomedical research. For its use, organization verification programs are provided, as the model has enhanced capabilities in sensitive areas.

Fable 5.2 created an OS prototype and formalized a mathematical proof

According to Anthropic, more than 90% of changes currently accepted into the company's main codebase are written by Claude under developer supervision. The new Fable 5.2 was also tested on large projects requiring coordination of multiple AI agents. In one experiment, the model managed a team of over 1,000 sub-agents who developed an operating system prototype over eight days.

The resulting system runs on a laptop and includes its own kernel, drivers, file storage, window interface, and installer. In another experiment, agents migrated a critical production service from Python to Rust. According to Anthropic, after the migration, the program's startup time decreased by approximately 26 times, and its reliability and efficiency improved.

A separate achievement was the formalization of Fermat's Last Theorem proof in Lean, a language that allows automatic verification of mathematical reasoning. A team of 1,000 Fable 5.2 agents completed the task in 28.7 hours with minimal human intervention. This was not about discovering a new proof but translating an existing mathematical result into a form accessible for computer verification.

At the start of the experiment, agents were provided with a set of intermediate mathematical statements and previously proven auxiliary lemmas. The team independently organized collaborative work and prepared a formal proof, which was seven times shorter than the result of a previous Anthropic experiment. At that time, a similar task took 11 days and required more active researcher involvement.

Models demonstrated results in physics, robotics, and cybersecurity

Anthropic also reported two results of Fable 5.2 in theoretical physics, related to black holes and holographic duality. The company claims that the model performed complex calculations necessary for investigating previously unsolved problems. The results are presented in scientific preprints, but further independent verification remains important for assessing the significance of these works.

In hardware engineering, Fable 5.2 designed three electronic devices: a home center with a voice assistant, a USB speakerphone, and a portable virtual pet. The model prepared technical specifications, selected components, designed printed circuit boards and enclosures, and wrote software. The physical manufacturing and assembly of the devices were carried out by Diode specialists with Claude's support.

In robotics, Fable 5.2 achieved a score of 64.9% on RoboDojo Sim, which tests an AI's ability to control a virtual bimanual robot based on camera images. The previous version, Fable 5.1, scored 27.6%. These figures are based on Anthropic's internal tests using a simulator and do not imply that the model demonstrates the same efficiency on any physical robots.

Meanwhile, Mythos 5.2 showed improvements in finding software vulnerabilities. In the internal CVE Rediscovery test, which included 1,829 previously known vulnerabilities, the model achieved a result of 89.6% compared to 78.3% for Mythos 5.1. According to Anthropic's calculations, the new version detected 10.2% more vulnerabilities than Claude Opus 5.5 and 14.4% more than Mythos 5.1.

One of the companies participating in the preliminary testing also reported that other tested advanced models missed about every sixth confirmed vulnerability found by Mythos 5.2. However, these results pertain to specific sets of software code and do not guarantee the detection of a similar share of errors in real projects.

How much do the new Claude models cost

Anthropic reported that Fable 5.2 and Mythos 5.2 generate responses up to 30% faster than the corresponding 5.1 versions. The company also reduced their usage cost. The standard price for both models is $8 per million input tokens and $40 per million output tokens.

Additionally, a Fast Mode is provided, increasing token generation speed by up to 2.5 times. Its cost is $16 per million input and $80 per million output tokens, which is twice the standard rate. Access to Mythos 5.2's accelerated mode is limited to organizations participating in Anthropic's verification programs.

Fable 5.2 is already available through Claude products and cloud platforms like Amazon Web Services, Google Cloud, and Microsoft Azure. To gain direct access to Mythos 5.2, specialists must pass verification in the Cyber Verification Program or Life Sciences Verification Program. Corporate clients can also use the model's capabilities through Claude Security for analyzing software code vulnerabilities.

Based on its own tests, Anthropic did not find significant degradation in the security metrics of the new models compared to previous generations. However, the company emphasizes that even an extensive set of tests cannot fully predict AI behavior after its implementation in real-world workflows.