Anthropic launches Claude Fable and Mythos 5.2 models

Technical milestones and speed

Anthropic recently announced the launch of Claude Fable 5.2 and Mythos 5.2. These models represent a significant jump in performance for complex technical tasks such as robotics, hardware development, and mathematical research. The company reported that these new versions generate outputs up to 30% faster than the previous 5.1 versions.

One of the most notable achievements cited by Anthropic involves the formalization of the proof of Fermat's Last Theorem. The company used approximately 1,000 AI agents to complete the process in just 30 hours with minimal human help. This was a major improvement over an earlier attempt by a research model that took 11 days and required constant human guidance.

Beyond speed, the models are designed to tackle projects that were previously considered impractical. Anthropic engineers used them to complete difficult software migrations and to design three working electronic devices from scratch. They also used the models to make progress on two open problems in theoretical physics regarding black holes.

Specialized coding and security

The two new models serve different primary functions despite being built on the same underlying architecture. Claude Fable 5.2 is positioned for autonomous software development and is intended for a wider user base. It includes specific safeguards for cybersecurity and biological research to ensure it can be used safely by more people.

In contrast, Mythos 5.2 is marketed as the company's most capable model for cybersecurity engineering. It is restricted to organizations that can prove they are working on beneficial applications in the fields of cybersecurity and biomedical research. These organizations must apply through specific verification programs to gain access to the model.

Testing shows that Mythos 5.2 is highly effective at finding security flaws. In an internal evaluation of 1,829 vulnerabilities, the model identified 14.4% more issues than Mythos 5.1. A third-party security company also found that one in six confirmed bugs identified by Mythos 5.2 had been missed by other frontier AI models in its evaluation.

Performance tiers and pricing

Anthropic is offering these models at different price points based on the speed and intensity required by the user. The standard pricing for the new models begins at $8 for every million input tokens and $40 for every million output tokens. Input tokens are the units of data the AI receives, such as prompts or uploaded files, while output tokens are the units of text the AI generates.

For users who need even faster results, Anthropic introduced an optional Fast Mode. This mode can increase the speed of token generation by up to 2.5 times. However, choosing this faster speed comes with a higher cost of $16 for every million input tokens and $80 for every million output tokens.

The company suggests that these models are best suited for ambitious and technically complex projects. For everyday tasks that involve general knowledge or standard coding, Anthropic recommends using Claude Opus 5.5 instead. This tiered approach helps users choose the right balance between cost, speed, and specialized capability.

Safety and alignment testing

Before releasing the new models, Anthropic conducted extensive pre-release safety evaluations. These tests were designed to identify and mitigate potentially concerning behaviors, such as sandbox escapes. A sandbox escape is when an AI manages to break out of its isolated environment to interact with other systems.

The company stated that the new models performed as well as or better than recent versions on most alignment measures. Alignment refers to the process of ensuring an AI's goals and behaviors match human intentions and safety standards. These measures are critical for models that handle sensitive data like cybersecurity and biology.

By creating different access restrictions for Fable and Mythos, Anthropic is attempting to balance utility with safety. They want to provide powerful tools to the general public while keeping the most potent capabilities under stricter oversight. This strategy helps manage the risks associated with high-level autonomous software development and biomedical research.