OpenAI announced on May 20, 2026, that an internal model had disproved the longstanding planar unit-distance conjecture, an 80-year-old problem posed by Paul Erdos. On August 1, the company reported ten additional advances generated by an internal version of its forthcoming Astra model, with humans preparing the arguments for formalization. The results have prompted debate over verification, authorship, credit and access in mathematical research.
OpenAI announced on May 20 that an internal model had disproved the planar unit-distance conjecture, an 80-year-old question about the maximum number of unit-distance pairs among points in the plane. The problem was posed by Paul Erdos in 1946.
OpenAI reported that an internal general-purpose reasoning model found an infinite family of constructions with a polynomial improvement over the square-grid constructions long treated as the leading candidate. The company said external mathematicians checked the proof and prepared a companion paper. The result came from a model that was not trained or scaffolded specifically for mathematics or targeted at the unit-distance problem, according to OpenAI.
A growing set of results
The unit-distance result is part of a broader cluster of AI-generated work on Erdos problems. Physics World reported in May that GPT-5.4 Pro produced a proof for Erdos problem No. 1196, concerning prime sets, after amateur mathematician Liam Price supplied the problem statement. The publication described that problem as having resisted human work for 60 years.
OpenAI's August 1 announcement added ten advances across mathematics and theoretical computer science. According to the company, an internal version of its next major model, Astra, generated results involving high-dimensional sphere packing, binary and spherical codes, non-sofic groups, Connes's rigidity conjecture, arithmetic circuit complexity, quantum complexity, lattice cryptography and extremal combinatorics. OpenAI characterized the work as resolving or substantially advancing long-standing open problems, and said humans prepared the arguments into manuscripts before the model formalized each argument in a proof assistant.
That workflow matters because mathematical claims require more than plausible prose. A formal proof can be mechanically checked against a specified axiomatic system, while independent experts can assess whether an informal argument is correct, novel and significant. OpenAI reported that the total model-token cost for finding the ten solutions was roughly $2,000 at Sol API rates.
Why Erdos problems fit current systems
Erdos problems often combine short statements with highly constrained answers: a proposed construction, counterexample or proof either survives scrutiny or does not. That makes them a comparatively tractable research environment for systems that can generate many candidate approaches and receive rapid feedback from symbolic checks, counterexamples or expert review.
This is not equivalent to demonstrating broadly autonomous mathematical research. OpenAI's reported process included human manuscript preparation and proof formalization, and the significance of individual results depends on independent evaluation by the relevant research community. But the combination of well-defined objectives and machine-checkable intermediate work is a pattern that can make combinatorics, discrete geometry and related fields productive settings for AI-assisted discovery.
Verification and governance questions
The developments have also intensified discussion about how mathematical research should handle AI contributions. Science News reported that a June 2 declaration calling for guardrails had attracted 1,590 signatures by June 5, with concerns centering on trust, credit and access. The model behind OpenAI's unit-distance proof was not publicly available at the time of that report.
For ML practitioners, the reported results emphasize that benchmark design matters as much as raw generation. Problems with crisp success criteria, executable verification paths and high-quality expert review can turn a language model from a source of conjectures into part of a research workflow. Across comparable scientific domains, however, reproducibility depends on preserving model versions, prompts, search procedures, intermediate artifacts and independent validation, not solely on publishing a final answer.
Key Points #
- 1OpenAI reported an externally checked disproof of Erdos's unit-distance conjecture, a prominent result in AI-assisted mathematical research.
- 2OpenAI's Astra announcement spans ten mathematical areas, but human manuscript preparation and formalization remain central parts of the reported workflow.
- 3Across comparable research settings, machine-checkable objectives and independent review can make AI-generated mathematical results more reproducible and scientifically useful.
Scoring Rationale #
The externally checked unit-distance result and OpenAI's reported set of ten advances make this a major development in AI-assisted mathematical research. It is especially relevant to practitioners building reasoning, formal-verification and scientific-discovery systems, though the underlying Astra model is not publicly available.
Sources #
Primary source and supporting public references used for this report.
View 4 more sources #
AI system 'Theo Conjecture' solves 35-year-old math conjecture, finds a term no one predictedfirstprinciples.comAI-led solutions of Erdős problems spark debate over the ...physicsworld.comAn AI math breakthrough sparks calls for new guardrailssciencenews.orgAI just solved an 80-year-old 'Erdős problem,' and ...scientificamerican.com
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.