One versus nine. That’s the ratio a T3 newsletter is reporting between OpenAI’s and Google DeepMind’s Erdős-problem achievements. According to The Rundown AI, Google DeepMind’s AlphaProof system, reportedly referred to internally as “AlphaProof Nexus”, has solved nine open problems from Paul Erdős’s unsolved conjecture set. That’s the claim. The supporting primary source isn’t accessible yet.
What’s confirmed: Google DeepMind runs an AlphaProof research program aimed at autonomous mathematical proof generation. That program uses formal verification combined with reinforcement learning, an architecture that’s been publicly detailed in prior DeepMind publications. The program’s existence and general methodology aren’t in question. What requires a primary source is the specific system name “AlphaProof Nexus” and the specific count of nine Erdős problems.
What’s also confirmed: OpenAI’s reasoning model autonomously disproved Erdős’s 1946 Unit-Distance conjecture, per prior analysis of that milestone published May 25. One Erdős conjecture. Independently documented. That’s the baseline the reported nine-problem figure is being compared against.
The 9-to-1 ratio is a journalistic framing from a single aggregator publication. It’s directionally useful if the underlying claim holds, but it shouldn’t be treated as a verified competitive score before a primary DeepMind source confirms the problem count.
The gap practitioners should care about: cost and compute efficiency. Reporting mentions the problems were solved at low compute cost, but specific figures aren’t accessible from a verifiable primary source and have been excluded from confirmed claims here. If cost efficiency at scale is a genuine characteristic of the system, it would distinguish this from prior high-compute mathematical AI results. That’s worth watching for when the technical paper becomes accessible, an arXiv preprint is reportedly pending publication.
What the four-month pattern in the registry makes clear is that AI mathematical reasoning has been advancing faster than most research teams’ expectations. The verified Erdős proof analysis from May 25 and the FrontierMath Tier 4 coverage from May 12 describe a trajectory, not isolated milestones. If the nine-problem claim holds up, it’s another data point on that trajectory, not a ceiling.
The catch is that “nine Erdős problems” isn’t a uniform category. Erdős posed thousands of conjectures at varying difficulty levels. The significance of nine solved problems depends entirely on which nine, at what difficulty tier, with what verification method. A primary source would answer those questions. The T3 reporting doesn’t. Research teams evaluating these results for practical application should wait for the arXiv paper before drawing conclusions about what this means for their specific use cases.
What to Watch
Don’t expect immediate applicability to most research workflows. Mathematical reasoning AI, even at this performance tier, is currently useful for a narrow class of formally stated problems. The tools require formal language environments, and most research questions aren’t already posed in that form. The value is in trajectory: what DeepMind and OpenAI are demonstrating about autonomous reasoning will propagate into more accessible tools over the next 12–18 months.
Wait for the primary DeepMind source before treating the nine-problem claim as confirmed, and wait for the arXiv paper before evaluating the architecture’s practical implications for your research stack.