GPT-5.6 Sol Ultra Mode: 64 Subagents on One Problem

GPT-5.6 Sol Ultra mode became the most-discussed AI release of the week when OpenAI published two PDFs on its own CDN on July 10, 2026: a three-page proof of the Cycle Double Cover Conjecture — an open graph-theory problem roughly 50 years old — and the two-page prompt that produced it, which authorized up to 64 concurrent subagents and a special configuration OpenAI calls multiagent v2.

The claim, if it survives review, would be a landmark: a frontier model producing a complete original proof of a long-open conjecture in under an hour. But two qualifiers are doing heavy lifting in that sentence. First, the proof is unreviewed — mathematicians expect verification to take days to weeks. Second, the 64-agent fan-out is not the Sol Ultra that shipped to customers on July 9; the default ultra configuration runs four parallel agents.

This guide separates what OpenAI actually published from what the coverage implies, extracts the genuinely reusable part — the orchestration playbook inside the prompt PDF — and works through the business question underneath the headline: what a 64-agent fan-out buys on hard problems, and where it is pure waste.

One day after GPT-5.6's Sol/Terra/Luna general-availability launch on July 9, two PDFs appeared on OpenAI's CDN: a three-page proof of the Cycle Double Cover Conjecture and the two-page prompt that produced it, both with document timestamps of July 10, 2026. There was no dedicated blog post on openai.com — the announcement came via an OpenAI staffer's post on X, with Ethan Knight quoted in OfficeChai's coverage as saying "We're excited to see what you all do with Ultra!"

The proof document makes the attribution itself. Its statement of AI use reads: "The proof in this note is entirely due to GPT 5.6 Sol Ultra and the writeup with Codex (with GPT 5.6 Sol)." That is OpenAI's own document, not a press paraphrase — the strongest anchor in this story. The conjecture it targets was posed independently by Tutte, Itai and Rodeh, Szekeres in 1973, and Seymour in 1979, which is where every outlet's "50-year-old" framing comes from.

In plain English, the conjecture: for every bridgeless graph — a network with no single edge whose removal disconnects it — there should exist a collection of cycles such that every edge appears in exactly two of them. The claimed proof reduces the general case to cubic graphs, applies the 8-flow theorem, then converts the flow labeling into a cycle double cover via what the document calls an elementary linear algebra argument. Per The Decoder's reporting, the run was provisioned for up to eight hours before "considering surrender" — and finished in under one.

Here is the distinction nearly every outlet blurred: 64 subagents is not what Sol Ultra does when you click "Ultra." Per OpenAI's developer documentation, the ultra mode that reached general availability on July 9 ships with 4 parallel agents by default on ChatGPT Work Pro/Enterprise and Codex Plus+, with the API equivalent described as a multi-agent beta on the Responses API. The Cycle Double Cover run layered a special configuration on top — the prompt explicitly grants "up to 64 concurrent agents available" under a mode it names multiagent v2.

That phrasing matters. "Available" means a ceiling, deliberately provisioned for one extremely hard, verifiable-output task — not a preview of the out-of-the-box experience. It is the same architecture direction OpenAI has been shipping since Codex's subagent general availability, scaled up 16-fold for a demonstration. Conflating the two leads to exactly the wrong takeaway for buyers: that a consumer subscription buys proof-an-open-conjecture capability.

Most coverage treats the proof as the story. For anyone who builds agentic systems, the two-page prompt is the more valuable document: it is a working specification for how to run a large fan-out on a hard, checkable problem without burning the budget on redundant work. Four rules stand out.

The fifth element is the one most amateur multi-agent setups skip: adversarial verification. The prompt mandates that every candidate sub-proof be checked against a specific list of failure modes — exact-two multiplicity, repeated-edge closed trails masquerading as cycles, parallel-edge 2-cycles, disconnected graphs, cutvertices, bridges introduced by reductions, and circular use of an equivalent restatement of the conjecture itself. Generation capacity is cheap in a fan-out; disciplined checking is the scarce resource, and OpenAI spent prompt real estate on it accordingly. If you have experimented with building custom subagents in Claude Code, this maps directly: the reviewer agent earns its tokens; the fifth parallel drafter usually does not.

OpenAI has not published a cost figure for the run. MLQ.ai's analyst estimates put the inference cost anywhere from roughly $275 to over $13,000 depending on assumed configuration and infrastructure — a 47-fold spread that itself tells you how little is disclosed. What is published: GPT-5.6 Sol's per-token pricing ($5 per million input tokens, $30 per million output, $0.50 per million cached input), and the simple arithmetic that concurrent agents multiply token consumption — N agents can burn up to N times the tokens of a single-agent run at the same per-token rate. The table below uses that formula for the ceiling column.

The middle row is the one that matters for buyers, and OpenAI's own GA benchmark tables published with the July 9 launch supply the honest context: on Terminal-Bench 2.1, Sol Ultra scores roughly 91.9% against plain Sol's 88.8%. A ~3-point gain from parallelization, on a benchmark already sitting in the high 80s — that is what fan-out buys on work the model can nearly saturate alone. The economics only invert when the problem is open-ended enough that most single attempts fail entirely.

As of this writing, the proof has not undergone formal peer review. Per MLQ.ai's reporting, professional verification is expected to take days to weeks, and existing formal-proof-assistant libraries — the Lean graph-theory corpus among them — are described as insufficient for research-level formalization of this result, so a machine-checked confirmation is not imminent either. Everything below the headline is therefore an OpenAI claim, and should be cited as one.

The expert commentary so far is genuinely mixed, and both halves deserve equal weight. In coverage carried by The Decoder and MLQ.ai, mathematician Thomas Bloom of the University of Manchester called it "a very nice proof" that is "short, elementary, and could have been discovered in the 1980s." But Bloom also criticized the document for omitting citation to a foundational 1983 paper by Bermond, Jackson, and Jaeger, whose ideas he says underlie the core reduction. "Praised for cleanliness, flagged for missing citations" is a very different status than "confirmed correct."

The conjecture's own history counsels patience. The Cycle Double Cover Conjecture has attracted a series of claimed proofs over the decades that later showed gaps or were withdrawn — a history of near-misses that is precisely why mathematicians will read every line before anyone updates the textbooks. And the survivorship question from the Hacker News thread still stands: a single published success, with no failure count alongside it, tells you the ceiling of what the system can do — not the rate at which it does it.

Strip away the graph theory and the CDC run is a case study in when parallel search is rational. The problem had three properties: most attempts were expected to fail (open problem), candidate answers were cheap to check mechanically (the prompt's adversarial list), and one success justified the whole budget. When those three hold, 64 agents is a search strategy. When they do not, it is 63 agents producing near-duplicate drafts of the same answer — variance without value. Our review of multi-agent orchestration patterns that actually work reaches the same conclusion from the practitioner side.

The forward projection worth making: expect the 4-agent default to creep upward and the 64-agent ceiling to become purchasable. The gap between "what the demo used" and "what customers get" is exactly the kind of gap vendors monetize, and OpenAI has already productized the pattern once — ultra mode itself is last year's research scaffolding sold as a toggle. When that tier arrives, the buying question will not be "how many agents" but "what is my verifier" — because the CDC prompt's real lesson is that OpenAI spent more instruction budget on checking than on generating.

For most businesses, the actionable content of this story is not the proof — it is the prompt's operating discipline, which scales down. A four-agent fan-out with an explicit approach registry, a rule against fixed assignment, and a mandatory adversarial check on every candidate output is implementable today on any frontier stack, and it is a materially better pattern than the naive "spawn five agents, take the best-looking answer" approach we see in most first-generation agentic builds.

The second implication is budget honesty. If analyst estimates for a single flagship run credibly span ~$275 to over $13,000, then fan-out costs are dominated by configuration choices you control — agent count, output length, verification depth — not by list price. Teams adopting multi-agent workflows should instrument per-run token accounting from day one, or the first genuinely hard problem will produce the first genuinely surprising invoice. This is the operating layer we build in our AI transformation engagements: routing rules for when a task justifies parallel agents, verifier design for the outputs that matter, and cost telemetry before scale, not after.

And the trend read: OpenAI publishing the prompt at all is the interesting move. It signals that orchestration technique — not raw model weights — is becoming the artifact vendors compete on publicly. The labs are converging on similar base capabilities; the differentiation is shifting to how you spend concurrent intelligence. That is a competition businesses can actually participate in, because prompts and orchestration patterns, unlike pre-training runs, are learnable by any team willing to study them.