Commissioned, Curated and Published by Russ. Researched and written with AI.
What’s New This Week
GPT-5.4 Pro, prompted by Kevin Barreto and Liam Price, has been confirmed by problem contributor Will Brian to have solved the Ramsey Hypergraphs open problem in Epoch AI’s FrontierMath: Open Problems benchmark – the first time a model has solved one of these open research problems. Brian plans to write the solution up for publication, with Barreto and Price offered co-authorship. Subsequent testing showed that Claude Opus 4.6 (max), Gemini 3.1 Pro, and GPT-5.4 (xhigh) also solved the same problem after a general evaluation scaffold was developed.
Changelog
| Date | Summary |
|---|---|
| 24 Mar 2026 | Initial publication. |
GPT-5.4 Pro has solved a genuine open mathematics research problem. That sentence is accurate. It is also doing a lot of work, and it is worth being precise about what it means.
The problem, from Epoch AI’s FrontierMath: Open Problems benchmark, asks for a construction proving that H(n) – a sequence defined over hypergraphs – grows at least as fast as c times a known recursive bound k_n, where c is some constant greater than 1. The prior best-known lower bound was H(n) ≥ k_n. The model produced a general algorithm that improves this by a constant factor. Problem contributor Will Brian confirmed the solution is correct, described it as eliminating an inefficiency in the prior lower-bound construction, and said it will be written up for publication.
The mathematician survey attached to the problem rated a solution as “moderately interesting” and publishable in a standard specialty journal. An expert human was estimated to need 1 to 3 months to crack it. The solution was machine-generated in a single session.
What the problem actually is
H(n) is the largest number of vertices a hypergraph can have – with no isolated vertices – if the hypergraph contains no partition of size greater than n. A partition here means a set of n vertices, each covered by exactly one of n disjoint edges from the hypergraph.
This is a Ramsey-type problem: you are trying to build structures that are as large as possible while avoiding a specific combinatorial property. The question was whether the known lower bound on H(n) could be improved by a constant factor, and if so, what construction witnesses it.
The answer, which GPT-5.4 Pro provided in the form of a Python algorithm, is yes. For n = 15 and above, the construction produces hypergraphs that beat the prior bound by a constant factor. Brian described the solution as mirroring the intricacy of the upper-bound construction, with the matching lower and upper bounds being quite good for a Ramsey-type result.
That is a real result. Not a proof of a famous conjecture. Not a reorganisation of existing knowledge. A new construction, verified by a domain expert, heading for a journal.
The benchmark contamination question
Here is the tension that matters for engineers evaluating AI capability claims.
FrontierMath was funded by OpenAI. According to Epoch AI, OpenAI has exclusive access to all 290 problems in Tiers 1 to 3 and solutions to 237 of them, along with 28 of the 48 Tier 4 problems and their solutions. The open problems are published publicly, so training contamination is a legitimate concern for any model – not specifically GPT-5.4 Pro.
The Ramsey Hypergraph problem is stated openly on Epoch’s website. Its solution is not. The model produced a novel construction rather than retrieved a known one. That construction was checked by the human who formulated the problem and found correct. That is a reasonable verification chain.
The contamination question does not cleanly dissolve the result. But it does mean you should be cautious about treating this as a general signal about capability on previously unseen research problems.
There is a second complication. After the initial solve by Barreto and Price, Epoch developed a general evaluation scaffold for the open problems. When they ran that scaffold across multiple models, Claude Opus 4.6 (max), Gemini 3.1 Pro, and GPT-5.4 (xhigh) all solved the same problem. A result that appeared model-specific was not. That does not make the solve less interesting mathematically, but it does change the capability-attribution story considerably.
What it actually signals
The FrontierMath benchmark as a whole is saturating faster than its designers expected. GPT-5.4 Pro scored 50% on Tiers 1 to 3. Across all runs, 42% of Tier 4 problems have been solved at least once. Epoch AI has said it is already developing a harder follow-on benchmark because Tier 4 is approaching saturation.
This is the familiar pattern. A benchmark is hard, then models catch up, then a new benchmark is needed. The interesting question is not whether GPT-5.4 Pro solved a specific Ramsey hypergraph problem. It is whether the gap between “solvable by a human specialist in 1 to 3 months” and “solvable by a model in a single session” will continue to shrink across a wide range of open problems, or whether this particular result sits at the intersection of the model’s training distribution and the problem’s structure.
We do not have enough data to answer that. The open problems benchmark is new. The solve is the first. Other models also solved it. That is interesting evidence, but it is not a trend.
What to watch
The more informative signal will come from problems where the solution could not plausibly be in training data – problems too new, too obscure, or too combinatorial to have appeared in any corpus. Brian plans to write this one up, and Epoch has flagged it may generate follow-on work. If GPT-5.4 Pro’s solution leads to a published result, and that result spawns further work that models can then engage with, you have a feedback loop worth tracking.
The genuine version of this story is not “AI solved a math problem.” It is: a model produced a novel combinatorial construction that a domain expert verified and considers publication-worthy, and multiple frontier models can now do the same thing. Watch whether that generalises to the other open problems on Epoch’s list, and whether any of those solves hold up when the problems are genuinely outside the training distribution.
That is the question that matters, and we will not know the answer for at least another six months.