OpenAI reasoning model claims to disprove 80-year-old discrete geometry conjecture

OpenAI announced on May 20 that one of its reasoning models produced a counterexample disproving a discrete geometry conjecture posed in 1946 and unsolved for nearly eight decades. The claim was published on OpenAI's research index and picked up by TechCrunch, which framed it carefully against the lab's prior embarrassing math episode where it overstated a result that working mathematicians quickly debunked. This time, OpenAI says the same mathematicians who exposed the earlier overreach have reviewed the new construction and back it up — a deliberate credibility move.

Mechanically, OpenAI describes the model working through a structured search guided by chain-of-thought reasoning, producing a candidate configuration that violates the conjecture's claim. The output is small and verifiable: rather than a long opaque proof, it's a concrete counterexample that human mathematicians can check by inspection. That verifiability is the point — it makes the result robust to the usual 'hallucinated proof' failure mode that has plagued LLM math claims, including OpenAI's own.

Competitively this lands in a busy week. Google's I/O keynote leaned hard on Gemini for Science (AlphaFold, AlphaGenome) and DeepMind's Co-Scientist hypothesis tool; Anthropic just hired Andrej Karpathy onto its pre-training team to push frontier research. OpenAI's geometry result is its own pitch for the same narrative — that scaled reasoning is now a research instrument, not just a chat product. HN's 371-point, 243-comment thread saw mathematicians weigh in more favorably than last time, though many still want to see the construction independently re-derived.

What to watch: whether the counterexample survives peer review in a math venue (not just an OpenAI post), and whether the technique generalizes beyond one conjecture. If reasoning models can reliably surface counterexamples to long-standing open problems, that's a qualitatively different capability claim than 'better benchmark scores' — and it would put pressure on Anthropic and Google to ship comparable demonstrations.