In July 2025, computers could solve only 3 of the 48 most hard math tasks of the FrontierMath Tier 4 test — these were tasks so hard that even good mathematicians needed many hours or days for them. In January 2026, artificial intelligence had already dealt with 17 specified tasks, and in March 2026, thanks to the fresh GPT-4.4 model, this figure increased to 20, or almost half of all tasks (42%).
And in September 2026, the fresh model, GPT-6 Astra, solved almost everything — from 97.6% to 98% of all the tasks of this test. It's like a student who, a year earlier, only took a single question on a hard exam, now passed it almost 100%.
Epoch AI, who created this test, considered it a signal that the test was exhausted — it became besides easy for the best AI, so it can no longer be verified who is truly the best. Thus, they decided to focus on something more difficult, on truly open mathematical problems, specified as those that have not yet been answered by anyone — humans or computers. This has an additional advantage: AI's tasks could not see earlier on the Internet, so the test better shows whether the model truly can think, alternatively than just associate facts that he has already read.
And here's a list of problems erstwhile opened that AI closed, so she solved:
🔘 Problem of Navier-Stokes equations
This is 1 of the 7 alleged Millennium Problems for which Clay's Math Institute has awarded a million dollars. These equations describe how liquids decision from water in the river to the air flowing around the wing of the aircraft. For almost 90 years, no 1 has known whether specified a smooth flow could abruptly break down and scope infinite velocity at 1 point. In September 2026, an interior yet unreleased OpenAI model (more capable than GPT-6 Astra) presented evidence that this was possible. Thousands of AI agents worked on the same problem in parallel for 88 hours. This is simply a very fresh and controversial consequence — mathematicians are just investigating it.
🔘 Non-sophical Group
Since 1999 mathematicians have wondered if there are certain mathematical structures (groups) that cannot be approached by simpler, finite arrangements. AI (preface version of Asta) constructed the first specified example.
🔘 Connes' Connes hypothesis overthrow
Since 1980, the mathematician Alain Connes assumed that certain groups could be clearly identified by a related algebraic object — as if by a unique fingerprint. AI showed that it was not actual to build 2 different groups with the same fingerprint.
🔘 Better packaging of spheres in many dimensions
It is simply a question of how to package the bullets as thick as possible (in many dimensions at once, not only in our three). AI improved the best known advanced limit of this density.
🔘 Better mistake correction codes
The point is how many safe messages can be stored in specified a way that a insignificant mistake does not turn 1 message into another — it is the basis for the operation of e.g. data transmission. AI gave much better numerical limits for specified codes.
🔘 advancement in calculating the matrix permanent
This is simply a hard computational problem with theoretical computing — AI has shown that any computational methods must by definition be slow (new lower limits of complexity).
🔘 Parallel repeatability in quantum games
A reasonably method consequence from quantum information explanation — AI proved any claim for a general version of a double-player game with quantum entanglement.
🔘 Problem of the nearest vector in the bars
This is an crucial issue for quantum computer-resistant cryptography — AI has shown that it is harder to solve than previously thought, what good news for the safety of specified encryption systems.
Ehrharta hypotheses
Concerning the volume of certain geometric lumps with peaks at the grating points, AI proved it.
🔘 Problem 183 Erdős (Ramsey's numbers)
Paul Erdős left hundreds of unresolved problems. This 1 concerns how many colors you request to usage to colour connections in the network to make certain a triangle of 1 colour appears. AI proved a new, much better lower limit.
Hadamard Matrix of 668
This is simply a peculiar table of numbers +1/-1 with circumstantial properties, used, among others, in coding theory. For this peculiar size it was not known whether specified a matrix existed. In August 2026, an Anthropic squad (three people + Claude model) reported the solution — Epoch AI, however, points out that this uncovering is temporary for the time being.
🔘 Better evidence for superpermutation
This is the shortest string of characters containing all possible shuffling of a given set of symbols. The GPT-5.6-Sol model improved the best known results for sets of 8, 9 and 10 symbols.
In most of these cases, AI generated an thought and a sketch of evidence, and people helped refine it and check it formally (e.g. in the thin program, which checks logical correctness step by step). So it is not a situation that AI sat down and did everything herself — alternatively a machine-manager collaboration in which the function of AI grows rapidly from period to month.
Grzegorz GPS Swiderski
]]>https://X.com/gps65]]>
]]>https://t.me/CanalBlogeraGPS]]>
