{"slug": "openai-has-solved-234-math-problems-anthropic-follows-with-57", "title": "OpenAI Has Solved 234 Math Problems, Anthropic Follows With 57", "summary": "OpenAI leads an informal scoreboard tracking AI-solved math problems with 234 attributed to its GPT and Codex models, while Anthropic's Claude follows at 57, according to the Vibe Mathed tracker. The cumulative growth curve shows OpenAI's count climbing sharply since late June, coinciding with the release of its Astra model, which alone accounted for ten new open problems at an inference cost of about $2,000. Of the 321 resolved entries, 63 percent were proofs and 37 percent were disproofs, with the longest-standing problem solved being the Jacobian conjecture at 87 years.", "body_md": "AI is now solving so many math problems that people have begun tracking data around these solutions.\n\nA site called [Vibe Mathed](https://x.com/haider1/status/2084308534465044902?s=20) has emerged as an informal scoreboard for the field, logging every open problem, conjecture, or long-standing puzzle that a frontier AI system has had a hand in resolving, and the running totals paint a fairly stark picture of where the credit is currently landing.\n\nOpenAI sits well out in front with 234 solved problems attributed to its GPT and Codex family, according to the tracker. Anthropic’s Claude models come in second at 57, ahead of a bucket labeled “agent systems / other” at 36, Harmonic’s Aristotle at 33, and Google DeepMind at 27. xAI’s Grok and the open-weights crowd, which includes DeepSeek and GLM, barely register on the chart with two and one solved problems respectively.\n\nThe cumulative growth curve on the tracker tells its own story. Through most of the winter and spring, OpenAI’s line climbed steadily but unremarkably alongside the pack, with Anthropic, Harmonic, and DeepMind bunched close together. Then sometime around late June, OpenAI’s count breaks away and starts climbing almost vertically, coinciding with the run-up to [Astra](https://officechai.com/ai/openai-says-it-has-solved-10-open-math-problems-using-astra-its-new-model/), the model OpenAI has been quietly showing to lawmakers in Washington. That single release reportedly accounted for ten new open problems in one batch, spanning sphere packing, group theory, and extremal combinatorics, at a combined inference cost the company pegged at around $2,000.\n\nAnthropic’s own line has a more gradual slope, but it has been picking up pace too. Some of that comes from Claude’s [Fable](https://officechai.com/ai/an-anthropic-researcher-says-fable-just-helped-him-disprove-the-85-year-old-jacobian-conjecture/) model, which an Anthropic mathematician credited with helping disprove the 85-year-old Jacobian conjecture, a result that touches the same list of century-defining problems as the Riemann Hypothesis. Anthropic’s [Mythos](https://officechai.com/ai/anthropic-says-that-mythos-has-also-solved-the-80-year-old-planar-unit-distance-problem-solved-by-openai/) model separately arrived at its own proof for the planar unit distance problem within days of OpenAI publishing one, apparently without any coordination between the two labs.\n\nA second chart on the tracker breaks the same 320-plus resolutions down by how much the AI actually contributed, rather than by which company gets the logo next to the result. “AI-discovered” problems, where the model is credited with the core insight, lead at 171. “AI co-developed” entries, where a human researcher worked alongside the model as something closer to a genuine collaborator, sit at 99. “AI-assisted” cases, where the model played a smaller supporting role, trail at 50. All three categories have accelerated sharply since March, which lines up with the period when labs began treating Lean formalization, the process of translating a proof into a machine-checkable format, as close to a baseline requirement for a credible math claim rather than a nice-to-have.\n\nThe tracker’s pie chart on outcomes shows 201 of the 321 resolved entries, or 63 percent, were proofs, meaning the AI system confirmed a conjecture was true. The remaining 120, or 37 percent, were disproofs, where the system found a counterexample that broke a standing assumption. That’s a meaningfully high disproof rate for a body of work this size, and it reflects something particular about how these systems have been used lately, hunting for the one counterexample that unravels decades of assumed truth rather than only grinding toward confirmations of what mathematicians already suspected.\n\nThere’s a separate leaderboard on the site worth a mention too: the problems that stood open the longest before finally falling. The Jacobian conjecture tops it at 87 years. The Unit Distance Conjecture, the one that drew both OpenAI’s and Anthropic’s models to the same result within days of each other, follows at 80. Rounding out the top five are the Pólya-Neumann balls problem and the “lost in a forest” gnomon question, both open for 72 and 70 years, and the Kac walk cutoff problem, also at 70.\n\nNone of this settles the argument mathematicians have been having since these claims started arriving, about whether a model surfacing a proof is the same thing as a mathematician understanding why the proof works, or whether formal verification alone is enough to call a result trustworthy. Harmonic’s Aristotle, for instance, first drew attention for [claiming an Erdős problem](https://officechai.com/ai/robinhood-ceo-vlad-tenevs-math-ai-startup-claims-to-have-solved-an-erdos-problem-that-was-open-for-30-years/) open for 30 years, only for the community to later note the specific variant solved was the easier of two the problem allowed for. What the numbers do settle is the pace. A field that measured its open problems in decades is now measuring them in a running weekly count on a public dashboard, and the gap between the lab in first place and everyone else is only getting wider.", "url": "https://wpnews.pro/news/openai-has-solved-234-math-problems-anthropic-follows-with-57", "canonical_source": "https://officechai.com/ai/openai-has-solved-234-math-problems-anthropic-follows-with-57/", "published_at": "2026-08-04 12:22:31+00:00", "updated_at": "2026-08-04 12:53:06.184638+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-products"], "entities": ["OpenAI", "Anthropic", "Google DeepMind", "xAI", "Harmonic", "DeepSeek", "GLM", "Vibe Mathed"], "alternates": {"html": "https://wpnews.pro/news/openai-has-solved-234-math-problems-anthropic-follows-with-57", "markdown": "https://wpnews.pro/news/openai-has-solved-234-math-problems-anthropic-follows-with-57.md", "text": "https://wpnews.pro/news/openai-has-solved-234-math-problems-anthropic-follows-with-57.txt", "jsonld": "https://wpnews.pro/news/openai-has-solved-234-math-problems-anthropic-follows-with-57.jsonld"}}