On May 20, 2026, OpenAI announced that an internal AI model had produced a counterexample to the unit distance problem—a conjecture Paul Erdős posed in 1946. It was the first historically significant proof ever generated by an AI, and it didn’t stop there. By August 1, the same unreleased model, Astra, had made ten additional mathematical advances, shaking the foundations of how we think about computation and creativity in mathematics.

What Happened

The unit distance problem asks: How many pairs of points in the plane can be exactly one unit apart? Erdős conjectured the answer grew roughly like n^(1 + c/log log n). For decades, mathematicians chipped away at the bounds, but no one had found a truly novel approach. Then OpenAI’s model did—it brought in techniques from a distant branch of math (combinatorial geometry interleaved with algebraic topology) that no human had successfully applied. The result wasn’t definitive: human mathematicians improved on it within weeks. But the novelty and the speed of subsequent breakthroughs (related techniques solved other Erdős problems within days) changed the conversation.

By August 1, Astra had solved ten more problems, including a new bound on the Erdős–Moser problem and a partial result on the discrepancy of sequences. The model was not publicly released, but OpenAI confirmed its architecture is a large-scale transformer trained on a curated corpus of mathematical literature and theorem-proving data. The key insight: AI can now not only verify proofs but find them—exploring search spaces humans would never consider.

Read the full announcement →

My Take

This is not just another “AI beats humans at game X” story. Erdős problems are the gold standard of mathematical depth—simple to state, fiendishly hard to prove. That an AI can produce a genuinely novel counterexample means we’ve crossed a threshold. The model didn’t brute-force check cases; it reasoned about structure. For developers and researchers, this signals that the next frontier of AI is not just language generation or code completion, but genuine discovery.

The implications for the math community are profound. We’re likely to see a wave of AI-assisted proofs, with models acting as collaborators rather than tools. The fear that AI will make mathematicians obsolete is overblown—human creativity and intuition still matter. But the skills that matter will shift: from “can I find a proof?” to “can I ask the right question and verify the AI’s output?”.

What to Watch

  • OpenAI’s next moves: Will they release Astra? If so, expect a flood of new results. If not, the gap between public and private AI capabilities will widen.
  • Human-AI collaboration in math: How will journals handle AI co-authors? The Erdős problem paper listed the model as a co-author—a first.
  • Security implications: The same reasoning ability that solves math problems can be turned to cryptographic analysis. Watch for AI-driven breakthroughs in breaking or strengthening encryption.