Nobody Can Run the Model That Broke Math, So Copy Its Org Chart

The most important thing OpenAI shipped this autumn wasn’t a proof. It was a staffing diagram.

I know that sounds dismissive of the math, and I don’t mean it that way. On September 21, 2026, OpenAI said an internal frontier model had resolved over 100 long-standing open problems, including Navier-Stokes existence and smoothness. Earlier, around September 9, the company described solving that problem after roughly 88 hours of compute using up to 10,000 coordinating AI agents. Then on October 6, 722 mathematical manuscripts landed on GitHub — 372 result families, supporting proof artifacts, Lean formalizations, and ten abridged summaries of how the model reasoned. Quanta ran a piece asking whether this is the end of math as we know it. Scientific American covered the drop the same day it happened.

And the model itself? Not released. Not named. Mathematicians and researchers are left holding artifacts and asking how to verify the claim of progress.

Why the number 10,000 matters more than 722

Here is where I’m writing from: I build bots. Multi-agent setups, tool-calling pipelines, the whole messy business of getting several language models to cooperate without talking themselves into circles. From that chair, “up to 10,000 coordinating AI agents” is the single most interesting number in this entire story.

Most agent architectures I’ve built or reviewed fall apart somewhere between five and fifty agents. The failure modes are boring and relentless:

• Context drift, where agent 14 is solving a slightly different problem than agent 3 because the handoff lost a constraint.
• Duplicate work, where a dozen agents independently attack the same subgoal and nobody notices.
• Consensus collapse, where agents agree with each other rather than with reality.
• Verification debt, where output accumulates faster than anything can check it.

Scaling to four orders of magnitude past my comfort zone means somebody solved coordination problems I actively fight every week. That’s the part I want the write-up on. Not the manuscripts.

The Lean detail is the tell

Of everything in the October 6 release, the Lean formalizations are what I keep coming back to. Lean is a proof assistant. It mechanically checks whether a proof holds. If you hand it a bad argument, it says no.

That’s a verification loop with teeth, and it’s the architectural pattern worth stealing. The reason 10,000 agents might not dissolve into noise is that there’s an external, non-negotiable judge in the system. Not a critic model. Not a confidence score. A checker that can’t be flattered.

Most bot builders, including me on lazy days, close the loop with another model. Agent writes, reviewer agent approves, done. It feels productive and it’s structurally weak — you’ve built a system that agrees with itself. If you’re running agents at any real scale, the question to ask is what your Lean is. A type checker. A test suite. A simulator. A database constraint. Something that returns a hard no.

What I’d actually change in my own stack

Three things, concretely, after reading through this news cycle:

• Stop treating verification as a post-processing step. If the checker runs only at the end, you’re paying for every wrong path in full. Put it inline.
• Separate result families from results. OpenAI grouped 722 manuscripts into 372 families. That’s an organizational choice that lets you dedupe and compare related outputs. Agent swarms generate heaps of near-duplicate work, and having a grouping key beats having a pile.
• Budget compute in wall-clock terms. Roughly 88 hours on one problem is a real figure a team signed off on. I usually cap agent runs at minutes because I’m scared of the bill. Some problems deserve days.

The withheld model is a design decision too

Publishing proofs while keeping the model private splits the output from the method. For a research field, that’s a problem — the field gets conclusions it can check but a process it can’t inspect or rerun. Reproducibility isn’t bureaucratic fussiness; it’s how claims become knowledge.

For those of us building smaller systems, it also means the useful lessons have to be inferred from the shape of the release rather than read from documentation. The ten abridged reasoning summaries are a thin window into something enormous. I’ll take them, but they’re breadcrumbs.

So the honest read: a field in shock, a GitHub repo of artifacts, and no model to examine. The math community will spend a long while checking the work. Bot builders should spend that time on the less glamorous question of how ten thousand agents stayed pointed in the same direction, because that answer is portable at any scale — including mine, and probably yours.

🕒 Published: October 7, 2026