On 10 July 2026, OpenAI claims that its GPT-5.6 Sol Ultra model produced a complete proof of a mathematical problem that had remained open for half a century, the Cycle Double Cover Conjecture. It reportedly did so in under an hour, orchestrating 64 AIs working in parallel. If the proof is confirmed, it is a historic milestone. The key word being: if.
There are announcements that are both dizzying and a call for caution. The one of 10 July 2026 is one of them. OpenAI has stated that its new AI had just solved, on its own, a problem that human mathematicians had failed to settle since the 1970s. Let's break down what actually happened, what it means, and why the mathematical community is holding its breath rather than applauding immediately.
The problem in question
The Cycle Double Cover Conjecture was posed independently by George Szekeres in 1973 and Paul Seymour in 1979. It belongs to graph theory, the branch of mathematics that studies networks of points (the vertices) connected by lines (the edges).
Its statement is simple to formulate, which makes its difficulty all the more striking. It asserts that in any "bridgeless" graph (a graph where removing a single edge never splits it into two separate pieces), one can always find a collection of cycles such that every edge is traversed exactly twice. Easy to state, extraordinarily difficult to prove: it is considered one of the most important open problems in graph theory.
The method: 64 AIs searching together
The most fascinating part may not be the result, but the method. OpenAI used the Ultra mode of its model, which splits a task among several sub-agents working in parallel, as we described when covering the release of GPT-5.6. Here, up to 64 sub-agents were launched simultaneously.
The principle is clever. In the early phases, the 64 agents are pushed to explore different avenues: varied algebraic angles, distinct formulations, independent structural inductions. The goal is to prevent the search from closing in too early on a single appealing but incomplete idea. It is a form of organised creative diversity, where the herd is prevented from running in the same direction. Then promising results are consolidated until they converge on a complete demonstration.
OpenAI published not only the proof, but also the prompt (the instruction) that generated it. Developers who examined it noticed something surprising: only about a fifth of the prompt concerns the mathematical problem itself. The other four-fifths serve to steer the model's behaviour, its perseverance, its ability to correct itself, its exploration strategy. In other words, knowing how to direct the AI was almost a bigger job than the maths problem itself. A valuable lesson on what the art of prompting really is in 2026.
Why caution is warranted
This is where enthusiasm needs to be tempered. An announcement is not a verification. The proof has not yet been reviewed and validated by the mathematical community (the peer review process), and this conjecture has a heavy history: it has attracted multiple false proofs over the decades, many of which seemed correct until a flaw was found.
The mathematician Thomas Bloom, who took a first look at the demonstration, called it "very pretty" and "elementary", estimating that it could have been discovered as early as the 1980s. But he also pointed out a flaw: the absence of citations of the foundational prior work on which it relies, an unusual shortcoming in serious mathematical work. Another limitation noted by experts: Ultra mode has its agents communicate in real time without producing an inspectable trace, which complicates understanding of the actual path taken to reach the result.
The deeper reflection
Beyond whether this particular proof holds up, this episode illustrates a shift worth reflecting on. What is changing is not just that an AI can do research mathematics, but that this capability is now accessible to almost everyone. The estimated cost of this demonstration would be somewhere between a few hundred and a few thousand dollars, a sum within reach of any reasonably funded team. Attacking a structured and difficult problem no longer requires a prestigious research institute.
The question this moment raises is dizzying and open-ended: what other hard, well-structured problems, in biology, cybersecurity, logistics, code analysis, will fall to 64 cooperative agents launched in parallel? The Cycle Double Cover Conjecture may have been just a first data point. But it also raises a subtler question about what it means to "understand" a proof: if a machine produces a correct demonstration that almost no one can follow step by step, have we truly gained in understanding, or only in answers? Mathematics has always been as much about the journey as the result. Perhaps that, more than speed, is what this moment invites us to rethink.