Another claimed AI math breakthrough arrives fast
The competitive cycle in frontier AI mathematics is accelerating. Just after OpenAI reportedly disproved the Erdős unit-distance conjecture, Anthropic employees are now saying Claude Mythos can solve the same problem as well, according to The Decoder.
The claim remains exactly that: a reported result described by Anthropic personnel and discussed publicly on X. That matters, because it places the story in a category increasingly common in advanced AI research: meaningful technical progress first circulated through labs, engineers, and working mathematicians before a full institutional paper or broader independent verification settles the question.
Even with that caveat, the reported development is significant. The Erdős unit-distance conjecture has been open since 1946. If multiple frontier systems can now find viable solution paths to a long-standing combinatorial-geometry problem, the relevant story is no longer a single headline-grabbing proof. It is the possibility that advanced models are beginning to show repeatable value on difficult research problems.
What Anthropic reportedly did
According to the source text, Anthropic used a test setup built after AI solved another Erdős problem. The system involved isolated Claude Code instances with Mythos access receiving the problem, exploring solution paths, and then passing summarized findings to other instances working independently. That detail is important because it shifts the conversation away from a single prompt and toward an agentic workflow.
In other words, the reported achievement is not being framed as a pure one-shot language-model answer. It is closer to a coordinated research harness: multiple model instances, problem decomposition, summarization, and iterative comparison of approaches. That makes the result less of a neat demo and more of a preview of how AI-assisted mathematical work may actually proceed in practice.
The source also says Mythos often took a different route than OpenAI’s model. If accurate, that suggests something even more interesting than duplication. Independent solution strategies are closer to genuine research value than simply reproducing one known line of reasoning.
Why the comparison matters
The article notes that mathematician Daniel Litt reportedly called Anthropic’s result “a bit worse” than OpenAI’s, while also saying Mythos found OpenAI’s solution too. That is a useful reminder that not all proofs are equal. In mathematics, elegance, compression, and conceptual novelty matter alongside correctness.
Still, the strategic takeaway is not that one lab’s proof was prettier than another’s. It is that multiple labs now appear to believe their systems can engage with open mathematical problems at a much higher level than earlier generations could. Once that becomes repeatable, the frontier moves from “Can an AI do this at all?” to “How often, how independently, and with what level of human oversight?”
The Decoder also mentions Google DeepMind’s recent announcement that an AI-assisted system solved nine Erdős problems using Lean, a formal proof language. That comparison sharpens an important distinction in current AI math work. Some systems rely heavily on formal-verification environments; others are being judged more on natural-language reasoning and agentic exploration. The field has not yet settled on which style is more revealing about raw capability.
The bigger shift
What makes this story durable is not only the specific conjecture. It is the speed of follow-on claims. Open problems in mathematics used to serve as clean markers of the boundary between human and machine reasoning. That boundary now looks more porous, especially when labs combine frontier models with orchestration tools that let them branch, compare, summarize, and retry.
There is still a large difference between a reported lab success and a stable, widely trusted research system. Verification, peer review, and reproducibility remain essential. But the pattern is hard to ignore: AI labs are no longer presenting only benchmark gains or polished consumer assistants. They are increasingly presenting systems as contributors to advanced knowledge work.
If these claims continue to hold up, AI math headlines will stop being rare anomalies and start looking like an emerging research category of their own.
- Anthropic employees say Claude Mythos can solve the Erdős unit-distance conjecture.
- The reported setup used multiple Claude Code instances rather than a simple one-shot prompt.
- The broader story is the fast pace of AI-assisted work on long-open math problems.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
