Anthropic releases Claude Fable 5.2 and Mythos 5.2 for complex technical work

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, two versions of the same underlying AI model aimed at demanding work in software engineering, cybersecurity, science, and hardware development. The company has also reduced the standard API price to $8 per million input tokens and $40 per million output tokens.

Fable 5.2 is broadly available through Anthropic’s platforms and major cloud providers. Mythos 5.2, which has fewer restrictions for sensitive cybersecurity and life sciences work, remains available only to organizations approved through Anthropic’s verification programs.

Anthropic says the new models generate responses up to 30% faster than their predecessors. An optional Fast mode can increase token generation speed by as much as 2.5 times, though it doubles the token price. The company continues to position Claude Opus 5.5 as its recommended model for most everyday coding and knowledge work. Fable is intended for longer, more ambiguous, and more difficult assignments.

Higher coding scores and agent coordination

Anthropic reports strong results on coding and computer-use benchmarks. On FrontierCode 2.0, an evaluation of changes that open-source maintainers would accept, Fable 5.2 reached 49.1% at its highest effort setting. Anthropic says that score exceeds results reported for Opus 5.5 and OpenAI’s GPT-6 Astra.

The company also says Fable 5.2 at a High effort setting performed roughly as well as Opus 5.5 at its maximum setting while costing about half as much per task in early FrontierCode testing. Such comparisons are based on Anthropic’s own testing and depend on the task, model settings, and amount of work required.

Fable 5.2 can serve as an adviser to smaller models or coordinate them as an orchestrator, according to Anthropic. In an internal experiment, the company says one model instance coordinated more than 1,000 subagents over eight days to create a prototype operating system. Anthropic says the result booted on a laptop and included storage, network and keyboard drivers, a graphical interface, and an installer.

For users, the important change is less about a single benchmark score than the model’s intended role. Anthropic is pitching Fable as a system that can break large assignments into smaller steps, review outputs, and continue working over many hours. That approach could be useful for large migrations, research projects, code audits, and data-heavy analysis, but it also requires careful human review.

Scientific claims and tighter access controls

Anthropic also describes a series of scientific and engineering demonstrations. It says Fable 5.2 helped solve two open questions in theoretical physics related to black holes and holographic duality. The company says the results appear in two preprints.

In mathematics, Anthropic says a team of 1,000 agents formally verified Fermat’s Last Theorem in the Lean proof language within 30 hours. This was a formalization of an existing proof, not the discovery of a new proof. The company also reports that Fable designed three electronic devices, including circuit boards, enclosures, and software, that worked after manufacturing without electrical modifications.

Mythos 5.2 is positioned as the more capable option for cybersecurity research. Anthropic says it finds and reproduces vulnerabilities more effectively than earlier versions. Because those capabilities can be misused, Fable includes safeguards that block many penetration-testing, exploit-generation, and research-biology requests. Anthropic says it reduced unnecessary blocks by about 55% compared with Fable 5.1 while retaining similar protections for high-risk work.

The company also says it has strengthened protections against model distillation, the practice of collecting outputs from an advanced model to train another system. For enterprise customers, Anthropic is expanding a private beta program that monitors patterns of model use within a customer’s own infrastructure, including in environments that use zero data retention.

Sources

• Claude Fable 5.2 and Mythos 5.2 – Anthropic
• Anthropic launches Fable 5.2 with a price cut and stronger coding scores – The New Stack