// news · research-papers · frontier-models2026-05-29source: deepmind / axios / hassabis

DeepMind AlphaProof Nexus solves nine open Erdős problems at low marginal cost — frontier-AI mathematics capability passes prior research-threshold benchmarks

DeepMind's AlphaProof Nexus solved nine open Erdős problems for the cost of a steak dinner — the most explicit recent demonstration that frontier AI mathematics capability has crossed historical research-threshold benchmarks. The result is what Demis Hassabis cited as empirical anchor for moving his AGI timeline to "a real possibility by 2029." The capability data point reshapes the mathematics-research conversation about AI's role in original proof discovery.

The mathematics-research substance is the consequential piece. Erdős problems are a canonical set of open mathematics problems posed by Paul Erdős that have served as benchmark-difficulty targets for proof techniques for decades. Solving nine open problems at low marginal cost is a fundamentally different capability data point from prior AI mathematics demonstrations (IMO grand-challenge problems, Olympiad-level problem solving) because it involves original proof discovery on questions that were intentionally selected to defy known techniques. The "cost of a steak dinner" framing is informally meaningful — it signals that the marginal compute required for the breakthroughs was low enough to be operationally trivial, not the multi-million-dollar training runs that the headline-grabbing prior demonstrations required.

The timeline-projection consequence connects the data point to the broader frontier-AI cycle. Hassabis explicitly cited the AlphaProof Nexus result as anchor for his "real possibility by 2029" AGI timeline. The methodology Hassabis cited applies more broadly: domain-specific capability uplift at low marginal compute cost is the operational signal that the underlying capability landscape is shifting faster than the headline benchmark numbers suggest. The Bayes-consistent orchestration framing is the parallel theoretical work that researchers will build on for the next generation of agentic systems built on top of the same underlying capability shift.

See our analysis →

Axios — Google plans win AI war AlphaProof Erdos → · DeepMind — AlphaProof Nexus mathematics research methodology → · HeyGoTrade — Google I/O 2026 DeepMind Talent Push →