Sam Altman previews the next generation of GPT, and 10,000 Agents have conquered the Millennium Prize Problems.

Sam Altman has revealed a key clue about the next-generation GPT.

On November 2 local time, at the ADIPEC 2026 Energy Conference held in Abu Dhabi, when talking about the Navier–Stokes mathematical breakthrough previously announced by OpenAI, Altman disclosed that the work was completed by 10,000 Agents powered by the upcoming next-generation GPT.

https://www.youtube.com/watch?v=2YWfVGpVtvc&t=332s

They worked continuously for about 4 days, operating in collaboration the entire time.

Altman specifically emphasized that this system has no traditional hierarchical structure and no unified leader. Each Agent can independently explore research directions, share progress, and decide what to do next based on existing results.

"No hierarchy, no boss, fully decentralized."

Altman believes that this represents a brand new way of working and conducting scientific research.

This statement also gives the public a clearer clue for the first time: the research model that has been used internally at OpenAI and far exceeds the capabilities of GPT-6 Astra is very likely to be the next-generation GPT to be released to the public.

10,000 Agents

How to Conquer the Century-Old Mathematical Dilemma?

The protagonist of this experiment is the Navier–Stokes equations.

They describe the motion laws of fluids such as water and air, and are one of the most basic mathematical tools in fluid mechanics. They are indispensable in fields ranging from aircraft design and weather forecasting to blood flow research.

However, a fundamental mathematical problem surrounding them has plagued researchers for nearly 90 years: in three-dimensional space, is it possible for an originally smooth fluid motion to suddenly produce a singularity in a finite period of time?

In simple terms, the fluid velocity becomes infinitely large in a finite time, and the originally smooth motion state gets out of control in the mathematical sense.

In 2000, the Clay Mathematics Institute included this problem in the seven Millennium Prize Problems, and set a $1 million bonus for each problem.

On September 8 this year, OpenAI officially announced this research achievement.

According to OpenAI, the internal model has constructed a set of mathematical proofs: under the condition of applying smooth external force, a fluid that is initially stationary and has a smooth motion state can form a singularity in a finite time while keeping the total energy finite.

The key to this construction is a vortex that continuously shrinks inward while being stretched. As the vortex core gets smaller and smaller, the rotation speed keeps increasing, eventually leading to velocity divergence within a finite time. OpenAI stated that this proof solves Cases C and D in the official formulation of the Navier–Stokes problem, and simultaneously released the paper and the Lean formal proof file.

What is really surprising is the research process behind it.

OpenAI disclosed that the entire experiment was launched on September 1. Researchers assigned different Agent groups to explore different mathematical problems and proof paths. The Agents could read cached Internet information and also run codes.

As the research progressed, the team found that breakthroughs in Euler equation-related problems provided new clues for Navier–Stokes research, so more computing resources were shifted to the latter. Subsequently, OpenAI used Codex to sort out the phased results of different Agent groups, and fed valuable discoveries back to other groups to promote follow-up research.

During the experiment, the internal model also underwent a further training upgrade, and the Agents were updated accordingly. In the end, the scale of Agents participating in the Navier–Stokes research reached about 10,000. They completed the core mathematical exploration in about 88 hours, during which 2.7 million messages were exchanged and about 130 billion Tokens were generated. After that, GPT-6 Astra spent another 17 hours completing formalization and verification work through Lean.

From problem decomposition and parallel exploration to result aggregation and mathematical verification, this is already a fairly complete AI-assisted scientific research process.

Next-Generation GPT

Core Changes May Lie in Agent Collaboration

The most noteworthy part of the information Altman revealed this time is that the capabilities of the next-generation GPT may no longer be limited to a single model completing a complex reasoning task.

This experiment demonstrates another way of working: letting a large number of Agents driven by the same base model explore in parallel, exchange research progress with each other, and gradually form research results that can be further advanced.

Of course, 10,000 Agents do not mean 10,000 independently trained super models. They are driven by the same set of internal models, and jointly complete tasks through tools, grouped communication and result integration. This system is not without cost.

In a podcast conversation with John Collison, co-founder of Stripe, on October 29, Altman revealed that OpenAI invested about 10 million US dollars in the Navier–Stokes research alone, and the figure may even be slightly higher than that. But he also made a bet: by September 2027, OpenAI is expected to re-complete the same work at about 1% of the current cost.

The cost reduction will come from smarter models and better research methods. This also means that the release of the next-generation GPT may bring two changes: the capabilities of the underlying model itself will continue to improve, and the efficiency of large-scale Agent collaboration will also be enhanced.

However, Altman has not yet announced the specific model number and official release time of the next-generation GPT, nor has he explained whether ordinary users can directly call the 10,000-level Agent collaboration system in this experiment.

Reference link: https://x.com/Hesamation/status/2117261058650112105

This article is from the WeChat official account "Synced", and is republished with authorization from 36Kr.