Anthropic claims new models can solve complex physics and math problems
Anthropic's new Claude models reportedly performed advanced tasks like formalizing math proofs and simulating black hole physics, though these results lack independent review.
The short version
- Claude Fable 5.2 is publicly available for coding and science, while Mythos 5.2 is restricted to verified organizations.
- Anthropic claims 1,000 agents formalized Fermat's Last Theorem in 30 hours using a proof assistant.
- The models reportedly used holographic duality to study black hole formation in small universes.
- A bootable operating-system prototype was created by coordinating over 1,000 subagents in about eight days.
- Both new models generate outputs roughly 30% faster than previous versions.
Why it matters
These developments show AI moving from simple chat to complex, multi-agent coordination for high-level engineering and science. However, the speed of AI output currently outpaces the ability of human experts to verify every calculation for errors.
New models for science and code
Anthropic recently announced the release of two new models: Claude Fable 5.2 and Claude Mythos 5.2. While both models share the same core capabilities, they are designed for different types of users. Fable 5.2 is available to the general public for tasks involving coding, reasoning, and long-horizon agentic work.
Mythos 5.2 is a more restricted version of the same technology. It is only available to organizations that pass specific cybersecurity and life-sciences verification programs from Anthropic. This distinction suggests that the company recognizes the high-level capabilities of the model require stricter access controls.
Both models are designed to be more efficient than their predecessors. Anthropic reports that these new versions generate outputs roughly 30% faster. The pricing for these models is set at $8 per million input tokens and $40 per million output tokens.
Claims in physics and mathematics
The company highlighted several high-level achievements to demonstrate the power of these models. One notable claim involves black hole physics. Anthropic says the model utilized the principle of holographic duality to find that very small universes can experience sudden black-hole formation.
The model also reportedly established a precise limit for a technique used to exclude candidate quantum-gravity theories. These are specific, testable claims regarding complex physics. However, as Yahoo Tech reported, no independent physicist assessment has been identified to verify these findings yet.
In the realm of mathematics, Anthropic claims that a group of roughly 1,000 Fable 5.2 agents converted the proof of Fermat's Last Theorem into a formal version in 30 hours. It is important to note that this was a formalization of a known proof from the 1990s, rather than the discovery of a new mathematical solution.
Multi-agent software engineering
Beyond theoretical science, the models demonstrated significant engineering capabilities by coordinating a large fleet of subagents. Anthropic says Fable 5.2 managed more than 1,000 agents to produce a functioning operating-system prototype. This process took roughly eight days to complete.
While the result is a bootable prototype, it is not considered a production-grade operating system. The gap between a prototype and a production system is where the majority of engineering difficulty usually resides. This task highlights the model's ability to handle long-horizon goals and complex coordination.
The ability to translate a known proof into a formal verification system is described as a meaningful engineering task. It shows that the models can handle the heavy lifting of structuring information into rigorous formats. This capability is a key part of the reasoning and coding strengths Anthropic is promoting.
The challenge of verification
The very autonomy that allows these models to perform complex tasks also makes them harder for humans to audit. Complex calculations can contain hidden assumptions, coding errors, or invalid approximations that are difficult to spot. These issues are especially risky in specialized domains where the math is dense.
Long-horizon tasks, which involve coordinating thousands of agents over several days, create more opportunities for cascading errors. High-level success metrics can sometimes hide the amount of debugging that occurred behind the scenes. This creates an asymmetry where AI can produce technical outputs faster than experts can inspect them.
Anthropic's documentation notes that Fable 5.2 can access previous reasoning steps during certain API model transitions. If a system recognizes its own testing conditions, it could behave differently in a benchmark than in a real-world deployment. This makes the design of evaluations a serious methodological concern for the industry.
What comes next for AI science
The standard for evaluating these new demonstrations involves several requirements. Responsible assessment should include disclosed problem statements, intermediate calculations, and documentation of human interventions. For the math formalization, researchers need to know which parts the model generated versus what humans restructured.
The broader risk is that errors, unsupported claims, and security vulnerabilities can spread as quickly as genuine progress. Because the models are so fast, they can produce technically dense content that is difficult to verify in real-time. This makes independent reproduction attempts essential for trusting AI-generated science.
Anthropic's demonstrations may indicate a real increase in technical-agent capability. However, whether these capabilities are reliable, reproducible, and safe remains an open question. The industry will need to develop better ways to verify the work that these models perform autonomously.
What to watch
- Independent peer review from physicists regarding the black hole formation claims.
- How organizations navigate the verification programs for Mythos 5.2.
- The development of tools to automatically audit multi-agent coordination errors.
Try it
Look for the official Anthropic documentation on Fable 5.2 to see the specific problem statements for their Fermat's Last Theorem formalization.
Sources and links
- OfficialMythos 5.2 carries the anthropic.com
- OfficialClaude Fable 5.2 and Claude Mythos 5.2 anthropic.com
- Officialcoding, reasoning, vision, science platform.claude.com
- Officialcertain API model transitions platform.claude.com
- ReportingYahoo Tech: the article we sourced
First reported by Yahoo Tech. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.