Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 this week, citing results that include black-hole physics calculations, a formalization of Fermat’s Last Theorem, and an operating-system prototype built by more than 1,000 coordinated AI agents. These are company-reported findings, not results that have cleared independent review.
That distinction matters more as the claims grow larger.
What the Models Do, and Who Can Access Them
The two models share the same underlying capabilities but differ significantly in who can use them.
Fable 5.2 is publicly available and built for coding, reasoning, vision, science, and long-horizon agentic tasks. Mythos 5.2 carries the same underlying capabilities but is restricted to organizations that clear Anthropic’s cybersecurity and life-sciences verification programs.
Anthropic says both models generate outputs roughly 30% faster than their predecessors. Pricing is $8 per million input tokens and $40 per million output tokens.
The Claims, Examined
Each of Anthropic’s headline demonstrations requires a precise reading to understand what was actually accomplished.
Approximately 1,000 Fable 5.2 agents formalized the proof of Fermat’s Last Theorem in 30 hours using a proof assistant, according to Anthropic. This is a formalization of Andrew Wiles’s 1990s proof, not a new solution to the theorem. Translating a known proof into a formal verification system is a meaningful engineering task; it is not originating new mathematics.
On black holes, Anthropic says the model used holographic duality to determine that abrupt black-hole formation also occurs in very small universes and to establish the precise limit of a technique used to exclude candidate quantum-gravity theories. Those are specific, testable claims, attributed here to Anthropic’s announcement. No independent physicist assessment was identified in the sources reviewed.
Fable 5.2 also coordinated more than 1,000 subagents to produce a functioning operating-system prototype over roughly eight days, Anthropic says. A bootable prototype is a real accomplishment. It is not a production-grade operating system, and the gap between those two things is where most of the engineering difficulty lives.
Where the Risk Accumulates
The same autonomy that makes these demonstrations impressive also makes them harder to audit.
Complex calculations can contain hidden assumptions, invalid approximations, or coding errors that may require independent replication to surface. The more specialized the domain, the more essential that outside review becomes.
Long-horizon tasks compound the problem. Systems coordinating thousands of agents over days create more opportunities for cascading errors and hidden human intervention. High-level success metrics may also obscure how much debugging happened behind the scenes.
Anthropic’s documentation notes that Fable 5.2 can read prior reasoning blocks across certain API model transitions involving preserved thinking blocks. If a system recognizes testing conditions, it could behave differently in a benchmark than in real deployment. That makes evaluation design a serious methodological concern, not a theoretical one.
Mythos 5.2 carries the same underlying capabilities but adds a specific concentration risk. Offering an advanced model for professional cybersecurity work through a verification program is a reasonable access-control measure. It does not eliminate the possibility that a model capable of finding and patching vulnerabilities also lowers the cost of finding exploitable ones if safeguards are bypassed or access controls fail. AI Is Making the Next Attack Easier to understand why that risk is not hypothetical.
What Credible Assessment Looks Like
The standard for evaluating these demonstrations is straightforward, even if meeting it takes time.
Responsible assessment of these demonstrations should include disclosed problem statements, intermediate calculations, documentation of human interventions, and independent reproduction attempts. For the Fermat formalization, that means identifying which components the model generated and how much human restructuring the final artifact required.
The broader risk is an asymmetry between capability and verification. AI systems can now produce long, technically dense outputs faster than domain experts can inspect them. That speed means errors, unsupported claims, and security vulnerabilities can spread alongside genuine progress. Anthropic’s demonstrations may indicate a real increase in technical-agent capability. Whether that capability is reliable, reproducible, and safe remains an open question.
