Epoch AI shifted FrontierMath focus toward an especially hard subset: 68 open problems associated with Paul Erdős. In runs reported in the September 2026 cycle, OpenAI’s pre-release GPT-6 Astra was the only model among five evaluated to post a nonzero score under the standard ~$300-budget configuration: about 3%, equal to two problems solved.
The two official credits are concrete and must be marked carefully. Astra disproved problem #74, at a reported cost of about $218 and roughly 15 hours of agentic work; and proved problem #126. Epoch AI’s Jaime Sevilla underscored the milestone: it is the first model to crack at least one item from this set. Rival models in the same standard bracket—GPT-5.6 Sol, GPT-5.5, Claude Fable 5.1, and Claude Fable 5—scored zero.
Verification: for the two official results, Lean / machine verification is reported. That matters more than the percentage. In AI-assisted mathematics, a “yes” without a verifiable artifact is advertising; a “yes” with a Lean-checked proof is an object the community can audit. Neural Pulse insists on not mixing the two standards.
Additional runs with looser rules and much larger budgets claimed progress on problems 1, 74, 126, 548, and 571, but those runs are not part of the official score and their aggregate cost exceeded $220,000. Treating them as the main benchmark would distort FrontierMath: the $300 protocol exists precisely to compare models in a reproducible, economically bounded regime.
The contrast between ~$218 for official #74 and >$220,000 in loose explorations illustrates an uncomfortable truth of 2026 mathematical AI: sometimes the finding appears “cheap” in a lucky run, and sometimes it emerges only after a compute burn no lab will publish as a fair product metric.
Do not invent Clay or Millennium prizes in this story. FrontierMath Erdős is not the Clay Institute list; these are Erdős-style problems, many open for decades, gathered to stress-test creative reasoning. Confusing prize brandings would be a serious editorial error.
For OpenAI, Astra supplies a positive headline amid a week heavy with Senate probing and regulatory pressure. A small but unique score on a set where every rival scored zero works as evidence of deep reasoning capability, distinct from coding benchmarks or offensive cyber tests.
For Epoch AI, the result validates the subset’s design: hard enough that the floor is zero, soluble enough that a frontier model can mark two. If the set were impossible, it would not discriminate; if it were easy, it would say nothing new after HLE and other saturated exams.
Credit norms: the math community will demand public Lean artifacts, human review of statements, and clarity about harness hints. An agent can “prove” something by mis-restating the problem. That is why official vs. looser runs is not pedantry: it is epistemological hygiene.
Neural Pulse recommends always citing: (1) official score ~3% / 2 problems; (2) #74 disproved and #126 proved; (3) rivals at zero; (4) Lean/machine-verified on the official ones; (5) extra claims off-score and very expensive. Any summary that omits (4) or (5) is biased.
It remains open how Epoch AI will handle future attempts to “buy” the leaderboard with extreme off-protocol budgets. If it does not firmly defend the $300 run as the only comparable figure, the Erdős set will dissolve into a race over who burns more money on sampling.
Computing historians will compare Astra’s 3% on Erdős with other “first nonzero” marks on saturating benchmarks: the value is in breaking the floor, not in declaring field dominance. Celebrating with sobriety protects Epoch AI’s and OpenAI’s credibility before an audience already skeptical from prior math and science hype.
Closing: Astra did not “solve Erdős.” It solved two items from a 68-problem set under protocol. That nuance fits an honest headline and separates Neural Pulse from social-media exaggeration. Progress is real; the myth of total automatic genius is not.
Next to watch is whether Claude Mythos, Gemini, or others match the score at the same budget, and whether Epoch publishes verification artifacts for external scrutiny. Until then, the official board remains: Astra 2, field 0.
