A Math Proof Went Out to the World With Zero Human Names on It

Here's the deal: look at the author line at the top of the proof. There's no person. It just says "GPT-5.6 Sol Ultra," with a note that Codex helped with the write-up. That's it. And that's the part that actually has the math community buzzing.

What OpenAI dropped on July 10 looks simple on the surface. Its top-tier configuration, "Sol Ultra," ran up to 64 subagents concurrently and produced a proof of the Cycle Double Cover Conjecture — a problem nobody had cracked in 50 years — in under an hour. OpenAI posted the full prompt and the full proof as PDFs on its CDN and basically said, "look for yourself."

The headline isn't "AI solved a math problem." That's happened before. The real story is two things. First, this isn't a computation or an answer-matching exercise — it's a general proof of an actual open conjecture nobody knew the answer to. Second, a human didn't hold the baton and feed hints. The model itself managed 64 subagents "aggressively and dynamically" to get there. Now, the fair caveat: this hasn't been peer-reviewed yet, and one mathematician has already flagged that it "skipped its citations." So this is still a claim, not a settled theorem. But it's a claim with real weight.

Meet the Players — OpenAI, GPT-5.6, and This Thing Called 'Sol Ultra'

Start with OpenAI. No introduction needed, but in this context the relevant fact is how hard the company has been leaning into "reasoning" and "agents." Where older GPT models aimed to answer smartly in one shot, the recent direction is: think long, try many branches, and self-verify. This proof is an extreme demo of exactly that direction.

Next, GPT-5.6. OpenAI shipped the GPT-5.6 lineup on July 9, split into three tiers. Sol is the flagship (highest capability), Terra is the lower-cost everyday-work option, and Luna is the fastest and cheapest. Pricing per 1M tokens: Sol at $5 input / $30 output, Terra at $2.50 / $15, and Luna at $1 / $6. It hit apps and the API at once, with the rollout reaching full global availability inside 24 hours.

Then the star of this story: Sol Ultra. This isn't a separate model — it's closer to a mode you attach to Sol. Its core trick is that it "coordinates multiple agents across parallel workstreams." The default is four agents. It burns a lot more tokens, but in exchange it delivers better results faster on demanding tasks. And Ultra isn't for everyone — it's gated to Pro and Enterprise in ChatGPT Work, and Plus-and-higher plans in Codex. For this proof, OpenAI pushed well past that default of four, all the way up to 64 subagents.

Finally, the man standing on the other side: Thomas Bloom, a mathematician at the University of Manchester. When OpenAI published the proof, he was among the first experts to seriously dig in, and his reaction is what keeps this story balanced. He called the proof itself "a very nice proof." But he also pointed out a decisive flaw. More on that below.

What It Proved, and How It Ran

First, the problem. The Cycle Double Cover Conjecture is an old chestnut in graph theory. The wording sounds intimidating, but the skeleton is simple. You've got a graph — dots (vertices) and lines (edges). Now consider a "bridgeless" graph: one where deleting any single edge won't split the graph into two pieces. The conjecture says that for any such graph, you can always find a collection of cycles that covers every edge exactly twice. Exactly twice. Not once, not three times — precisely two.

George Szekeres posed it in 1973, and Paul Seymour posed it independently in 1979. Over the next 50 years, plenty of mathematicians took a run at it, and along the way several "I proved it" attempts collapsed once errors surfaced. It's a problem full of traps. That's the key point — it's famous not because it's hard to state, but because it's easy to get wrong.

So how did Sol Ultra approach it? The prompt is public, so we can peek at the method. The instructions went like this. Manage up to 64 subagents "aggressively and dynamically." Early on, maximize diversity — have each agent chase a different formulation, a different algebraic angle, a different structural induction, independently, so you don't converge on one direction prematurely. Mark agents stuck in dead ends as "blocked," and only revisit them if a genuinely new mechanism shows up. Separately, run adversarial agents whose job is to hunt for traps — repeated-edge trails, disconnected graphs, cutvertices, circular reasoning that smuggles in an unproven equivalent. And finally: keep working for at least eight hours before you even consider giving up.

The result? It was handed an eight-hour budget and finished in under one. The proof document went up as a PDF on OpenAI's CDN on July 10, with the mathematics credited entirely to Sol Ultra and Codex noted as assisting with the write-up. OpenAI's Ethan Knight quipped, "We're excited to see what you all do with Ultra!"

Item Detail
Problem Cycle Double Cover Conjecture (every bridgeless graph has a set of cycles covering each edge exactly twice)
First posed Szekeres 1973 · Seymour 1979 (independently)
Model used GPT-5.6 Sol Ultra (parallel-agent mode)
Subagents Up to 64 running concurrently
Compute budget Instructed "at least 8 hours" → finished in under 1 hour
Published July 10, 2026
Authorship GPT-5.6 Sol Ultra (no human), Codex assisting on write-up
Verification status Pre-peer-review · still a "claim"

Who Gains What

For OpenAI the win is obvious: the "autonomous research agent" narrative. The frontier-lab race has moved past a few benchmark points and onto "can our model actually produce new knowledge?" A proof of a 50-year-old open problem is a perfect billboard for that narrative. It also justifies the existence of Ultra as a premium paid tier — "this is why it costs more."

There's an upside for mathematicians and researchers too. Read coldly, this isn't AI replacing proofs wholesale; it's AI settling in as a tool that sweeps the search space fast. Even Bloom described the proof as "short, elementary, and could have been discovered in the 1980s." Paradoxically, that means a machine found — via parallel search — a "simple path" humans may have walked right past. For researchers, that opens up real use as an idea generator and a counterexample hunter.

For developers and companies, the upside is more practical. The real product here isn't a math proof — it's the orchestration pattern that produced it. Running 64 agents through a flow of maintain-diversity → block dead ends → adversarially verify → converge. That template ports straight over to code refactoring, bug hunting, and complex research, not just math. The key is that Ultra offers it in a productized form.

Past Cases — Wins and Flops

This picture isn't brand new. In 2023, DeepMind's FunSearch put an LLM in an iterative search loop and found new constructions in combinatorics (the cap-set problem). A win — but that was closer to "finding better examples" than "proving." In 2024, DeepMind's AlphaProof and AlphaGeometry solved International Math Olympiad problems at a silver-medal level, which made waves — but those were exam problems that already had answers. This is a different flavor entirely: Sol Ultra touched an open problem humanity didn't know the answer to.

To be fair, you have to flag the flops and the controversies too. There have been plenty of AI-generated "proofs" that looked plausible but broke at a decisive step. And the Cycle Double Cover Conjecture specifically has a track record of "we proved it" claims collapsing under scrutiny over the past half-century. So this time, both the press and mathematicians are drawing a careful line: this is a claim awaiting verification, not a done deal.

And Bloom's critique lands right on that spot. While crediting the proof's completeness, he criticized the paper for not citing foundational 1983 work by Bermond, Jackson, and Jaeger. In his words: "This is a frequent issue with AI-generated proofs and papers: they use ideas and proof strategies taken from the literature without proper citation." Why does that matter? Missing citations aren't just a courtesy problem — they blur the line between "this idea is genuinely new" and "this rediscovered something that already existed."

How Competitors Counter

Google DeepMind is the first counter that comes to mind. DeepMind's AlphaProof lineage leans hard into formal verification — using proof assistants like Lean so a machine can check the argument line by line. Where OpenAI's proof is narrative-first and requires a human to read and judge it, the DeepMind camp can play the "ours is machine-verifiable" card. Against formal verification, even a citation-gap controversy becomes far easier to defend.

Anthropic comes at it from another angle. The Claude line has branded itself around "reasoning reliably over long horizons" and "agent safety." If Anthropic responds, it's likely to differentiate less through flashy conjecture-cracking and more through "our agent is verifiable and honest about its own limits." In fact, this citation-gap flap could balloon into a safety-and-transparency issue — "the AI doesn't know where it got its ideas" — and that's a crack Anthropic is well-positioned to pry open.

The open-weight camp — DeepSeek, Alibaba's Qwen, and the like — can mount yet another counter. "64-agent parallel orchestration" is ultimately a "just run a lot of it" approach, and that isn't the exclusive property of closed frontier models. Expect follow-on attempts to stack an orchestration framework on top of open-weight models and reproduce similar multi-agent search cheaply. Whether Ultra's real moat is the model or the orchestration gets put to the test right here.

So the core competitive question is this: is Sol Ultra's feat something only this model can do, or something that reproduces given a good-enough model plus good orchestration? If it's the former, OpenAI's moat runs deep. If it's the latter, similar demos will pour out from all directions within months. For now, too early to call.

So What Actually Changes

For developers — the immediately usable lesson is "don't try to one-shot it with a single agent." OpenAI just demonstrated that on hard problems, what works is running multiple diverse agents in parallel, mixing in adversarial agents that deliberately hunt for counterexamples, and pruning dead ends fast. Whether you use Ultra or build your own orchestration, this pattern transplants straight into code and research work.

For investors — this is a different kind of signal from benchmark-score inflation. OpenAI answered the "can AI produce new knowledge?" question with a concrete case. But it's still at the "claim" stage, so don't over-read it. Whether this is the starting gun for a huge R&D-automation market or a single well-staged demo depends on how well follow-on cases reproduce. For now, it's a watchlist-and-observe situation.

For companies — it's a signal that a slice of R&D and expert knowledge work could shift to "agent orchestration." That said, as the citation-gap flap shows, verifying the provenance and internal consistency of the output is still a human job. The thing to do right now is draft a list of "which exploratory tasks in our domain we could hand to AI agents." This method is strong on tasks where you search broadly for good candidates — much stronger than on tasks with a single fixed answer.

For everyday users — honestly, you won't feel much change right away. The everyday questions you throw at ChatGPT won't suddenly transform because of this. But it matters. It's a signal that AI is at the doorstep of moving from "a tool that organizes existing knowledge" to "a tool that creates knowledge that didn't exist." If that direction is real, the very pace of discovery in science, medicine, and engineering could change within a few years.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Almost no direct impact. Your ChatGPT doesn't suddenly get smarter today. But if you do a lot of research, coding, or complex analysis, take it as a hint that "multi-agent parallel search" is going to become the standard, replacing the "one big prompt" habit.

— Is this actually settled, or still just a claim? Still a claim. The full proof is public, but it hasn't been peer-reviewed, and the Cycle Double Cover Conjecture has a history of "proofs" collapsing. Bloom liked the proof itself but flagged the missing citations. Independent verification by the math community is still needed before it counts as a theorem.

— Is this really ahead of the competition? Too early to call. DeepMind's strength is formal verification (a machine checking each line), so it can counter with "ours is verifiable," and the open-weight camp can imitate the orchestration cheaply. Whether Ultra's real moat is the model or the "just run a lot of it" method will be decided by how many reproductions show up over the coming months.

Sources

Numbers are as of announcement and may change.