“We called for pacing, not pausing,” Dianne Penn, Anthropic's head of product management for research, told CNBC.
Fable 5.2 and Mythos 5.2 share the same underlying model but differ in their safety restrictions and intended audiences. Fable includes safeguards that limit potentially dangerous cybersecurity and biological research activities, allowing Anthropic to offer it broadly. Mythos provides more extensive capabilities in those areas and remains restricted to approved organizations participating in Anthropic's verification programs.
Both models cost $8 per million input tokens and $40 per million output tokens. Anthropic also introduced an optional faster processing mode that can increase token generation speed by up to 2.5 times, with pricing of $16 per million input tokens and $80 per million output tokens.
The company positions Fable 5.2 as a specialized option for difficult technical problems rather than a replacement for its everyday Claude models. Anthropic continues to recommend Claude Opus 5.5 for most routine programming and knowledge work, with Fable intended for more demanding assignments involving engineering, mathematics, and scientific research.
“These are just net new things that are completely novel, and so they’re frontier use cases,” Penn said. “We expect that over time to maybe trickle to the Opus class, but really right now we do see a difference in terms of performance that we think those frontier use cases benefit from.”
Anthropic's published evaluations show substantial improvements across several technical benchmarks. On FrontierCode 2.0, which measures whether AI-generated software changes meet the standards of open-source maintainers, Fable 5.2 achieved 49.1% at its highest reasoning setting. That compares with 45.5% for Opus 5.5 and 42.8% for OpenAI's GPT-6 Astra.
The company also reported a 69.9% score on FrontierSWE v2, an evaluation of autonomous software engineering, compared with 56.3% for Fable 5.1.
Anthropic highlighted several examples of the model's capabilities during internal testing. In one experiment, Fable 5.2 coordinated more than 1,000 AI subagents over eight days to develop a functioning prototype operating system. The resulting software could boot on a laptop and included features such as file storage, networking, windows, and a login interface.
The company also reported advances in scientific and engineering applications. According to Anthropic, Fable 5.2 helped resolve two previously open theoretical physics problems involving black holes, while a coordinated group of approximately 1,000 agents completed a formal computer-verifiable proof of Fermat's Last Theorem in 30 hours.
In hardware testing conducted with circuit-board manufacturer Diode, the model designed three functioning electronic devices, including a voice-assistant hub, a USB speakerphone, and a handheld virtual pet. Anthropic said all three worked without requiring electrical modifications.
Mythos 5.2, meanwhile, focuses particularly on advanced cybersecurity research. Anthropic said the model identified more previously documented software vulnerabilities than earlier versions and demonstrated improvements in reproducing and repairing security flaws.
In an internal evaluation involving 1,829 known vulnerabilities, Mythos 5.2 outperformed Mythos 5.1 and Opus 5.5 across the reasoning settings tested. Anthropic said its capabilities are available to verified cybersecurity organizations, while enterprise customers can also access Mythos-powered vulnerability scanning through Claude Security.
The technical improvements arrive alongside additional safety measures. Anthropic said it conducted extensive evaluations before release, including testing by external organizations such as the UK AI Security Institute, METR, and Frontier Design.
The company examined potentially unsafe autonomous behavior, attempts to bypass restricted computing environments, and risks associated with advanced cybersecurity capabilities. Anthropic said the models generally performed at least as well as recent Claude systems on alignment measures, although it acknowledged that existing evaluations cannot fully predict how advanced AI systems will behave after deployment.
Anthropic also disclosed that it deliberately limited certain capabilities that could accelerate AI development itself. During training, the company restricted improvements in areas including accelerator optimization and AI model architecture research, while seeking to preserve progress in conventional software engineering and scientific applications.
The decision reflects Amodei's earlier argument that frontier AI development should proceed with greater attention to safety risks, even as Anthropic continues releasing more capable systems.
That approach is taking place against the backdrop of an anticipated public offering. Anthropic is reportedly targeting a valuation between $1 trillion and $1.5 trillion, below the $2 trillion figure discussed in recent months.
The company has reported substantial financial growth ahead of the potential listing. In an amended prospectus, Anthropic disclosed $33.24 billion in revenue for the nine months ending September 30, an increase of approximately 1,200% from the corresponding period a year earlier.
Anthropic's safety strategy is also a consideration for prospective investors. The company's prospectus outlines risks associated with developing increasingly advanced AI systems while operating in a competitive market that rewards rapid improvements in performance.
Claude Fable 5.2 is generally available through Anthropic's platforms and supported cloud providers, including Amazon Web Services, Google Cloud, and Microsoft Azure. Access to Mythos 5.2 remains limited to qualified cybersecurity and life sciences organizations, with additional availability through Claude Security for enterprise customers.
The launch gives Anthropic more capable models for specialized technical work while maintaining separate access rules for sensitive applications. With its IPO preparations advancing, the company is seeking to demonstrate that it can continue improving frontier AI performance without abandoning the safety restrictions it has publicly emphasized.
This analysis is based on reporting from Anthropic & CNBC.
Image courtesy of Anthropic.
This article was generated with AI assistance and reviewed for accuracy and quality.
