Anthropic Model Claimed to Crack 18-Year-Old Quantum Conjecture in 3 Days

The claim is specific: an unreleased Anthropic model reportedly resolved a quantum computing conjecture that had stood for 18 years, and did it in three days. That is the entirety of what has been disclosed. No benchmark name, no paper, no coauthor, no independent verification. For a practitioner, a conjecture's age is not evidence of difficulty calibration, and three days is not evidence of generality. The number being reported, 18 years, measures how long a specific open problem sat unresolved in a field, not how hard it was relative to other unsolved problems or how much compute, prompting, or human steering went into the three-day timeline.

Anthropic has not released the model, which means there is no way to run the proof claim against a second instance, a different prompt set, or a skeptical mathematician's checklist. Mathematical proofs claimed by automated systems have a history of looking complete and then failing on a single lemma a referee catches months later. The reported event has not gone through that referee process. Nothing here states whether the proof was checked by a human expert in the specific subfield, whether it was formally verified by proof-checking software, or whether Anthropic itself has made any internal claim beyond what has circulated. Anthropic did not disclose a verification method, and that gap matters more than the headline number.

Practitioners deciding whether to act on this should separate three questions that are currently collapsed into one claim. First, did a model produce a complete, correct proof, which requires peer or formal verification that has not been reported. Second, did it do so with minimal human scaffolding, which requires knowing the prompting and iteration process, also undisclosed. Third, is the result reproducible on a comparable problem, which cannot be tested because the model is unreleased. Anyone using this as a signal for procurement, research partnership, or competitive positioning is currently relying on a single unverified claim about an inaccessible system, not a benchmark result with a method section.

The broader pattern is familiar from prior claims of automated theorem discovery: an impressive headline precedes the slower, less reported process of mathematicians checking the work line by line, sometimes finding gaps that take the claim from solved to promising. An 18-year-old conjecture did not become open because mathematicians were careless, and a three-day solution, if real, deserves the same scrutiny any human-authored proof would get before entering the literature. Until Anthropic publishes the model, the proof, or third-party verification, the responsible position is that a claim exists, not that a conjecture has been closed.

What would change the calculus is simple to name and currently absent: a published proof object, a named verifier, and access to the model or its outputs for independent testing. Until one of those three shows up, the three-day figure describes a report about a report, and the 18-year figure describes the problem's age, not its resolution.