Anthropic releases Claude Fable 5.2 and Mythos 5.2 for complex engineering

New models for heavy lifting

Anthropic has upgraded its high-capability models, Claude Fable and Claude Mythos, to versions 5.2. These models are specifically built to handle the most difficult technical challenges that standard AI might struggle with. They show measurable improvements in reasoning, coding, and vision capabilities compared to previous iterations.

While Fable and Mythos share the same underlying model, they serve different purposes based on safety requirements. Fable 5.2 includes strong cybersecurity and biology safeguards, making it available to all users for general high-level work. Mythos 5.2 is designed for specialized use cases like cybersecurity engineering and requires application through specific verification programs.

The new versions are also more efficient for developers to use. They generate outputs up to 30% faster than the 5.1 versions and are priced at $8 per million input tokens and $40 per million output tokens. Users can also opt for a fast mode that increases generation speed by up to 2.5 times for a higher price point.

Automating software and hardware

Anthropic reports that Fable 5.2 has already been used to complete high-value software projects that were previously considered too slow or costly. For example, a team of agents migrated a critical production service from Python to Rust, resulting in a service that starts about 26 times faster. The model also demonstrated the ability to fix stubborn bugs in minutes and optimize performance across various services.

The model's capabilities extend into physical hardware engineering as well. In a collaboration with Diode, Fable 5.2 designed three working electronic devices, including a home voice-assistant hub and a USB speakerphone. It wrote the specifications, selected components, routed the four-layer circuit boards, and modeled the 3D-printed enclosures with minimal human input.

Beyond software and hardware, the model showed significant progress in robotics. It outperformed other frontier models on the RoboDojo benchmark, which tests how AI controls a simulated two-arm robot using camera images. It successfully completed tasks like solving a puzzle cube and unscrewing a bottle cap, even though it was not explicitly trained on those specific actions.

Advancing science and math

Fable 5.2 has demonstrated the ability to solve complex problems in theoretical physics. It autonomously completed a calculation regarding black hole formation in very small universes, a task that was laid out in 2003 but deemed too laborious for humans to finish. It also found the exact limit of a technique used for over 15 years to rule out theories of quantum gravity.

In the field of biology, Mythos 5.2 helped researchers tackle protein folding for gene therapies. It found a way to accurately predict the structure of a whole capsid—the protein shell of a virus—by writing code to run existing models on overlapping protein chains. In testing, it accurately modeled 28 out of 34 attempted capsids, both natural and engineered.

The models also excel at mathematical formalization, which is the process of translating a proof into a language a computer can check. A team of 1,000 Fable 5.2 agents fully formalized Fermat’s Last Theorem in just 30 hours. This was significantly faster than a previous research version that took 11 days and produced a proof that was seven times shorter.

Cybersecurity and safety

Claude Mythos 5.2 is positioned as the most capable model for cybersecurity engineering. It can catch more vulnerabilities than previous models and is better at reproducing and fixing them. One security company found that one in six confirmed bugs found by Mythos 5.2 were missed by all other frontier models they tested.

Anthropic implemented these models with a focus on safety and alignment. They ran extensive prerelease testing, including novel evaluations for specific behaviors like sandbox escapes. The company believes that high-stakes risks from misalignment remain low, and the models performed as well as or better than recent versions on most alignment measures.

Fable 5.2's safeguards are designed to produce fewer unnecessary blocks while remaining robust against misuse. Additionally, Anthropic implemented their strongest-ever protections against distillation attacks, which are methods used to extract knowledge from a large model into a smaller one. Full details on these safety measures are available in the official system card.

How to use these models

Anthropic suggests that users should not start with Fable or Mythos for every task. For most standard coding and knowledge work, they recommend starting with Claude Opus 5.5. Users can then move upward to Fable or Mythos as the scope and complexity of the work increase, or downward to Sonnet or Haiku for simpler needs.

Fable 5.2 can also function in specific roles like Advisor or Orchestrator. In an Advisor role, smaller models like Haiku or Sonnet can consult Fable to help solve harder problems. As an Orchestrator, Fable can delegate simpler tasks to smaller models while it handles the high-level planning and coordination.

The models are designed for ambitious, open-ended technical work where abstract reasoning is required. Early testers have found that Fable 5.2 handles ambiguous, data-heavy tasks better than previous models. It is particularly effective for research and analysis tasks that require complex design and judgment.