GPT-5.6 closed a 30-year-old convex optimization gap in 148 minutes, sort of
Around July 17, 2026, an interesting math story hit the front page of Hacker News. Someone used GPT-5.6 Sol to prove a lower bound in convex optimization that had been open since roughly the mid-1990s. The proof took about 148 minutes to produce. The thread is worth reading carefully, because the headline and the substance diverge in ways that matter.
Here is the claim. The problem class is convex, Lipschitz functions over a bounded, spherical domain. The question is how many function evaluations any algorithm needs in the worst case, in d dimensions. The proof establishes a lower bound of Omega(d^2) function evaluations. That matches the upper bound of an existing algorithm from roughly 30 years ago, which closes the gap. Showing upper bounds on time complexity is the easy direction: it is just the runtime of your algorithm. Lower bounds are usually the hard direction, because they require constraining every possible algorithm, not just the one you wrote.
I want to be honest about what is and is not impressive here, because the HN thread already did most of that work for me and I would rather pass it through than inflate it.
The upper bound algorithm dates to the mid-1990s. The matching lower bound had been missing since.
Wall time of the final GPT-5.6 Sol session that produced the proof.
The same author spent roughly a year trying with GPT-5.4 and GPT-5.5 before Sol landed it.
The function class was named in the prompt. The author and Sol co-wrote the prompt itself.
What the headline leaves out
The most useful correction in the thread came from someone who dug into the author's own account of how this happened. The author had spent about a year working on this problem with GPT-5.4 and GPT-5.5. That work, including failed attempts and prior constructions the author had discussed with earlier models, fed into the final session. So the honest framing is "a year plus 148 minutes" rather than "148 minutes flat."
There is a second caveat that is arguably larger. The prompt itself included the technique that solved the problem. As the author described it, they threw all the reasonable approaches at the model without ranking which one to explore most, and those approaches would all be reasonable to a person who knows the field. The function class that actually worked, max of affine functions, was in the prompt. The author also used GPT-5.6 to help write the prompt itself, after giving Sol their previous work, some ideas, a clear problem description, and the CDC prompt (referencing the cycle double cover proof from earlier in July) as a template.
So when you read "GPT-5.6 used a prompt to close the gap," that is technically accurate and also technically misleading. The prompt carried the candidate construction. The model carried the execution. Sorting out exactly how much credit goes to the human versus the model is not something the thread resolves, and I do not think it resolves cleanly either.
One thing the author openly wondered about
Whether Sol leaned on chat memory of those earlier sessions. The author noted the main construction that worked is not very different from something they had discussed with GPT-5.5 previously. They are not sure how much that memory carried over. Neither am I. It would not shock me if it did, and it would not shock me if it did not. Memory inside a closed model is exactly the kind of thing that is hard to audit from outside.
Why the math itself is real, even if the framing is generous
The most credible expert comment in the thread comes from someone who says they know this field. Their read: the conjecture is more niche than the cycle double cover conjecture that OpenAI's team claimed a week earlier, but it is a real contribution. A lower bound of Omega(d^2) function evaluations for optimizing a convex Lipschitz function over a spherical domain. The restriction to a sphere is not a real restriction; you can do a change of variables for any bounded domain.
My read of why this is harder than it sounds: the matching upper bound comes from a 30-year-old algorithm. Lower bounds on time complexity are nasty because you do not get to argue from a single algorithm that works. You have to argue that no algorithm in the class can do better. That means constructing an adversarial instance, typically a max of affine functions style construction, that forces any algorithm to spend at least d^2 function evaluations to be sure where the optimum is. The function-class construction is doing the heavy lifting, and the prompt more or less handed the model that family.
A second comment in the thread connects this to machine learning in a way that got pushed back on, and I think the pushback is mostly right. The original commenter claimed that optimizing a convex bounded Lipschitz function is exactly what most modern statistical learning models are based on. The correction is that modern neural network objectives are notoriously nonconvex, catastrophically so from the point of view of classical optimization theory. The honest version is that core optimizers like Adam and stochastic gradient descent are lifted from convex optimization theory, and near minimizers functions tend to be locally convex, but the deeper theorems of convex optimization, including the bound above, do not transfer to nonconvex settings in a clean way. So this is a real result about a real class. It is not a new constraint on how neural network training behaves.
The context this sits in
This is the second GPT-5.6 math claim in roughly ten days. The first was the cycle double cover conjecture. OpenAI released a three-page proposed proof on July 10, 2026 alongside the two-page prompt used to produce it. The prompt was given to GPT-5.6 Sol Ultra, instructing it to use up to 64 concurrent agents, employ adversarial agents throughout to check candidate proofs, and spend at least 8 hours. They also released a formalization in Lean. The cycle double cover conjecture was posed in the 1970s independently by Tutte, Itai and Rodeh, Szekeres, and Seymour, and asks whether every bridgeless undirected graph admits a collection of cycles such that each edge is contained in exactly two of them. It is one of the better-known open problems in graph theory.
GPT-5.6 itself launched on July 8, 2026 per CNBC. The Sol tier is the flagship. The launch announcement cites state-of-the-art results across coding, knowledge work, cybersecurity, and science benchmarks, with Sol scoring 80 on the Artificial Analysis Coding Agent Index and 53.6 on Agents' Last Exam. None of that is what makes the math work possible, but it is the substrate the math sat on. The number to watch for the math claims specifically is the reasoning budget. Ultra coordinates four agents in parallel by default and trades higher token use for stronger results on demanding tasks, which lines up with how the cycle double cover proof was set up.
The wrinkle about prompts and credit
The thread spends a lot of time on a question that I do not think has a clean answer: if a person writes a 10-page prompt that contains the candidate construction, and a model fills in the proof under that prompt, who did the work. One commenter's framing, which I find useful, is that the prompt itself is largely AI-generated. The author used GPT-5.6 to help write the prompt. That involved giving the model access to prior work, then iterating on the phrasing. So the partition is not "human specifies, model executes" in a clean way. It is more like a longer thinking process where the human and the model took turns doing the part each was better at, and the model's part coincided with the part that gets called "the proof."
Another commenter pointed out the thing I keep coming back to. In math, a proof is a proof. We do not know in advance whether we can get there, and getting there is the hard part. The route matters less than the destination. But a different commenter pushed on exactly that: how much of mathematical research is actually "grab a conjecture off the shelf and pump until it falls." Some of it is. Most of it is not. A lot of the work goes into asking which question is worth asking, into theory-building where the definitions are the contribution rather than the theorems, and into the aesthetics of how something is proved. Models are not doing that part yet. The "proof at any cost" loop is the part models are good at, and it is the part that is easiest for outsiders to mistake for the whole field.
Validation, and why that is where the ground shifts
One commenter made a point about proof that I think is the structural reason these announcements are not just hype. A proof can be validated as either correct or wrong. That kind of hard feedback is brutal on the optimism and shortcuts that models tend to drift toward in domains where validation is softer. If a model produces a wrong proof, a Lean formalization runs it down and refuses to typecheck. If a model produces a wrong code refactor, the test suite may pass for the wrong reason and the bug ships and gets discovered weeks later by a user. Math has the harder ground truth.
That does not mean every LLM-generated proof is good. It means the rejection signal is honest, which is more than you get in most software. The cycle double cover proof came with a Lean formalization. I did not see a Lean artifact attached to the convex optimization result in the thread, and I would want one before taking the claim fully to the bank. A 10-page prompt plus a model's chunk of reasoning is a different epistemic status from a proof that has been typechecked by a mechanical checker. The former can still be right. The latter is treated as right.
What I actually take from this
Two things. First, the bare headline "GPT-5.6 closed a 30-year gap in 148 minutes" is the wrong frame even though it is the frame that traveled. The honest version is that a person who had been grinding this problem with prior GPT releases for a year, who knew the field well enough to enumerate the reasonable constructions, produced a prompt with the most likely construction already in it, and Sol finally pushed the proof across in a 148-minute session. That is a different and I think more interesting story than the headline. It is a story about a human and a model doing combined work where neither side alone was enough for the prior year.
Second, the part of math that models are good at is not the same part that most working mathematicians spend their time on. Pumping a well-specified conjecture until it falls is real work and it is hard, but it is a narrow slice of the field. When you lower the cost of that slice to roughly "a year of expert prompting plus a couple of hours of compute," the field does not collapse, but the things people choose to spend their time on probably shift. Some of the people in the thread who actually do this work sound shaken. Some sound dismissive. I do not know which reaction ages better. I do think the people who treat this as "an obvious frontier the field has to reckon with" are closer to right than the people who treat it as "low-hanging fruit that does not count," but I would not put it more strongly than that.
If you want to read this for yourself, the HN thread is item 48957779 from around July 17, 2026. The author's account of how the proof happened is in the comments, including the year of prior work and that the working construction was in the prompt. The cycle double cover release and its Lean formalization are on OpenAI's site from July 10, 2026. Wikipedia's cycle double cover entry already references both. Read the source accounts before you decide how impressed to be. I found I was less impressed than the headline wanted and more impressed than the dismissive comments wanted, with the gap mostly explained by "a year plus 148 minutes."