GPT-5.6 Sol Ultra has produced a rigorous proof of a mathematical conjecture that remained open for fifty years. A mathematician verified the proof and found it correct.

It is yet another result of this kind in just a few days. On July 12 the model had closed the Cycle Double Cover Conjecture using 64 subroutines in parallel. On July 18 it solved a convex optimization problem, with a benchmark showing that agent architecture mattered as much as raw model power. Now another conjecture, another human verification.

Why it matters to you. Repetition is the point. One lucky proof can be called coincidence. Multiple verified results on different problems signal that something has shifted in how frontier models construct novel reasoning on unexplored ground. If you use AI for analysis or decisions where no pre-written answer exists anywhere, this is the signal that the capacity to reason beyond the known has become reproducible.

It does not mean every output is correct. A paper published the same day on arXiv shows that models still violate basic properties of statistical consistency when fragmented into subpopulations: they know things they don’t propagate reliably in aggregate estimates. The distance between “can prove a theorem” and “can analyze without error” remains large. The first is a domain where correctness checks out binary. The second doesn’t.

In detail

The series of mathematical results from GPT-5.6 Sol Ultra follows a pattern worth understanding for anyone using AI in their work.

The first result (July 12). The Cycle Double Cover Conjecture is a graph theory problem open since the Seventies: every graph without bridges admits a cyclic cover of its edges where each edge appears exactly twice. GPT-5.6 Sol Ultra attacked it with 64 subroutines in parallel, each exploring a branch of the proof. The mathematician who verified the result described it as elementary and correct (two words that in mathematics mean more than a thousand adjectives).

The second (July 18). A convex optimization problem, with one important difference from the first: an independent benchmark showed that agent architecture mattered as much as the model itself. The control loop, delegation to subroutines, and cross-verification made the difference between a plausible answer and a proof that holds.

This result (July 16). A new conjecture open for fifty years, again with human verification. We don’t yet know the details of the proof or the specific name of the conjecture, but the fact that a mathematician checked it and confirmed it is the central point.

Why mathematics is a special case. Theorems have a formal structure a model can explore systematically, with immediate feedback: a step is correct or it isn’t, and an expert can verify it. This is different from saying AI reasons in general. It’s more precise: in domains where correctness is verifiable binary, frontier models can construct novel proof paths that no one had written before.

The counterpoint. The paper by Wolf and colleagues on arXiv (July 16) documents the flip side. The researchers tested whether models satisfy the law of total probability: if you ask a model to estimate a probability over a population, then ask it to estimate the same thing over subgroups and aggregate, the results should match. They don’t. Models possess knowledge of subpopulations but don’t reliably propagate it in aggregate estimates. The researchers call this phenomenon macro fallacy: estimates reconstructed from subgroups often align better with human data than direct estimates, but the model doesn’t know it has this information and doesn’t use it when you ask for a general estimate.

For anyone using AI in their work, the consequence is concrete. The capacity to construct rigorous reasoning exists, but it is fragmented. The same model that proves a theorem can botch basic statistical aggregation. If your task has a possible formal verification (mathematics, code, logic), AI can surprise you. If verification is fuzzy (estimates, qualitative analysis, judgment), more caution and more cross-checks are needed.

What remains to verify. We don’t have the full text of the proof, nor do we know which conjecture was closed. Verification comes from a professional mathematician, but we don’t know the details of the review process. The sources gathered provide theoretical context (the link between compression and intelligence discussed by Karpathy, statistical inconsistencies documented in Wolf’s paper) but don’t directly confirm the event. As always with frontier results, it’s worth waiting for formal publication before drawing firm conclusions about general implications.