AI Research• 10 min read

September 2026: AI Agents Become Researchers, and Navier–Stokes Tests the Rules

On September 8, OpenAI said about 10,000 AI agents working for 88 hours had proved that the forced three-dimensional Navier–Stokes equations can blow up in finite time, a result aimed at one of mathematics' Millennium Prize Problems. Within days the claim faced questions about verification, credit and data provenance, while Anthropic reported that Claude now leads 26% of its own AI research. September was the month AI agents went from doing tasks to producing research, and the research world started writing the rules.

Author: Macaulan Chiaramonte

September 2026 put numbers on agentic science. OpenAI's swarm of about 10,000 agents produced a proof of finite-time blow-up for the forced Navier–Stokes equations in 88 hours, then checked it in Lean. Nine days later, Anthropic said Claude leads 26% of its AI research and development, up from under 1% in February.

The reaction mattered as much as the results. The Clay Mathematics Institute said the problem had "apparently been settled" but promised a "deliberately unhurried" review. Twenty-five Fields Medalists warned that the goals of AI companies and of mathematicians are misaligned, and a priority dispute put the provenance of customer data at the center of the story.

Ten Thousand Agents, Eighty-Eight Hours

OpenAI's announcement described a research system rather than a single model. According to VentureBeat, an internal model described as significantly more capable than Astra, which OpenAI had begun releasing five days earlier, ran as roughly 10,000 concurrent agents. The agents could run code, consult a cached version of the internet and communicate with other agents in their group.

The timeline was compressed. VentureBeat reported that the agent run began on September 1, after rumors of a rival team's progress, and reached its result on September 5, about 88 hours later. Astra then spent roughly 17 more hours formalizing the proof in Lean, the language that lets a computer check every logical step.

The scale was industrial. The Navier–Stokes work alone consumed about 130 billion output tokens, which VentureBeat estimated at roughly $6.5 million at retail API prices; Quanta Magazine put the compute at several million dollars. The internal model is not available to ChatGPT or API customers, and OpenAI said it does not intend to claim the $1 million prize.

This is a different kind of result from August's Astra math announcement. Astra worked through ten problems for about $2,000 in total; the Navier–Stokes run organized thousands of agents into a coordinated research group and spent more than a thousand times as much. That suggests the unit of scientific capability is shifting from the model to the orchestrated swarm.

What Was Proved, and What Was Not

The Navier–Stokes equations describe how fluids move. Whether their solutions can break down is one of the Clay Mathematics Institute's Millennium Prize Problems, six of which were still unsolved before September, each with a $1 million prize. OpenAI described its result as the "existence of forced blowup in R^3 and T^3," covering both unbounded space and the periodic case.

The claim is narrower than many headlines suggested. OpenAI's result shows that a smooth three-dimensional fluid pushed by a smooth external force can reach infinite velocity in finite time. The unforced version, the harder case most mathematicians pursue, remains open.

Quanta still reported the result as solving the Millennium Problem, and the author of its official description agreed. Princeton's Charles Fefferman, who wrote the Clay Institute's problem description, told Quanta, "I was thrilled that the problem was solved." He also credited earlier human work, naming mathematicians Córdoba and Martínez-Zoroa as "the heroes of the story."

Verification is moving on two tracks. The Lean formalization is strong machine-checked evidence that the proof is internally correct, but Quanta noted that humans still have to confirm the formal statement matches the problem as posed. On September 11, the Clay Mathematics Institute said the problem had "apparently been settled" and that its evaluation would be "deliberately unhurried."

Clay's rules require publication, a two-year wait and acceptance by the mathematical community before a prize is considered. For a result produced in 88 hours, the gap between machine speed and institutional verification is the real story, and it applies well beyond mathematics.

The Provenance Question

The announcement landed in the middle of a dispute. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had been independently pursuing the same problem using both OpenAI's Codex and Anthropic's Claude. Buckmaster publicly questioned whether their work had informed OpenAI's system.

Buckmaster was careful about the limits of what he knew: "I do not know whether our data was used. I am not accusing anyone of anything." OpenAI denied using their work. "We did not use their prompt or proofs to prompt our models," OpenAI's Sébastien Bubeck said at a press briefing, according to Scientific American.

The sentence that matters most for enterprises came from OpenAI's initial statement: "we cannot rule out that de-identified data derived from their usage of our products helped improve our models." OpenAI later said Buckmaster's recent Codex prompts "could not have influenced the system in any way." The Batch noted that the update did not address whether earlier sessions affected the model.

No public evidence settles the question either way, and this article does not try to. The structural issue stands on its own. When a vendor's general models improve from aggregate usage, the line between a customer's private research and the vendor's next breakthrough is set by contract terms and data pipelines that customers often cannot inspect.

De-identification removes who did the work, not what the work contained. For an individual chatting with an assistant that distinction may not matter; for a research team whose unpublished method is the asset, it is the whole question.

Fields Medalists Push Back on Speed

Three days after the announcement, 25 Fields Medalists published a declaration titled "A Severe Misalignment of AI in Mathematics", posted by Terence Tao. It names no company or problem, but its central claim is blunt: "The goals of the AI companies and the goals of the mathematical community are severely misaligned."

The signatories argue that solving problems has always been a means to understanding, not the end. Their specific complaint is haste: solutions "announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others."

They also frame it as a wider problem. In their words, "years of training have traditionally served not only to produce a final answer or product, but also to develop understanding and the ability to formulate new questions." Mass-produced answers, they warn, can break the chain by which mathematicians learn from one another, and errors can spread without adequate checking.

The declaration does not set out demands. It says the issues must be addressed urgently by the mathematical community, by the companies developing the technology and by society more broadly.

The critique transfers directly to corporate R&D. An answer that arrives without an explanation, a reusable method or credit for the work it built on is hard to trust, hard to extend and hard to defend in a patent or publication dispute. Speed is a capability; legibility is what makes it usable.

Inside the Lab: Claude Leads 26% of Anthropic's AI R&D

While mathematicians argued about one proof, Anthropic measured something broader. On September 17 it published a prototype index of how much of its own AI research Claude performs, rating tasks on an automation scale from AL0 to AL5 developed by Epoch AI.

The headline: as of August, Claude "leads" 26% of Anthropic's AI R&D work, up from under 1% in February. Leading means completing most of a task end-to-end from a high-level prompt while a human supervises. More than 90% of the work now runs at the collaboration level or higher.

Anthropic was explicit about the limits: "Claude is not operating fully autonomously for any measured subset of AI R&D work." The ratings rely partly on judge models, which landed within one level of human ratings 97% of the time. The Associated Press noted it was unclear from the disclosure how close Anthropic believes it is to recursive self-improvement.

The oversight numbers are as notable as the automation ones:

  • Scale: about 30,000 research and engineering agents running at the same time inside Anthropic
  • Coverage: 100% of agent actions pass through real-time monitors before they execute
  • Intervention: only about 0.002% of decisions, roughly 1 in 47,000, are blocked
  • Review: about one or two transcripts per 1,000 are flagged for offline human review
  • Safety compute: in a measured July week, about 6% of AI R&D compute, and about 12% of AI-driven R&D compute, went to safety

Monitoring every action at that scale is the lesson August taught, applied in production: autonomy is only deployable when supervision scales with it.

A Different Template: 950 Agents and a New Enzyme System

Anthropic's September 23 announcement showed a slower, more conventional path from agent output to scientific claim. About 950 Claude agents spent 21 hours and 210 million tokens searching biological data before one spotted a previously uncharacterized system in bacteriophages.

The system, which Anthropic calls array-associated reverse transcriptases, pairs a reverse transcriptase enzyme and a partner gene with a DNA repeat array resembling those of CRISPR. According to Anthropic, it has "a set of characteristics that have only ever been found together in a handful of other systems."

The division of labor was explicit. Humans wrote the initial prompt; Claude agents searched databases, generated hypotheses and interpreted data; and "all of the lab work is performed by human scientists." Its 210 million tokens are a small fraction of the 130 billion output tokens in OpenAI's Navier–Stokes run, a reminder that agent science does not have to start at swarm scale.

Anthropic was modest about the result. It released a preprint, said "we don't yet know its function," and presented the work as early findings. CRISPR pioneer Feng Zhang called it "genuinely intriguing" and said it "merits further investigation." The episode pairs agent speed with human verification and a candid account of what remains unknown.

What Enterprise and Technology Leaders Should Take From September

Few organizations will attack Millennium Problems, but many are starting to point agent swarms at proprietary research, code and data. September's events translate into concrete practices:

  • Audit data-use terms: confirm whether usage data, including de-identified or derived data, can improve a vendor's general models, and get the answer in writing
  • Budget for swarms: one research run at OpenAI's scale cost millions in compute, so agent research needs the same portfolio discipline as other R&D
  • Plan verification first: formal checks, test suites and lab validation should be designed before the agents run, not after
  • Record provenance and credit: log which people, prompts, datasets and prior work fed each result, so ownership questions can be answered later
  • Measure automation honestly: an AL-style scale, rated by people as well as models, shows where agents lead and where humans still drive
  • Monitor every action: Anthropic's pre-execution monitoring of all agent actions is a sensible bar for agents with access to sensitive systems

The common thread is legibility. The organizations that benefit most from agentic research will be the ones that can show what their agents did, what they built on and how each result was checked.

Also in September

  • OpenAI ships Astra: on September 3, OpenAI began a phased rollout with its Daybreak cybersecurity program, with paid ChatGPT plans and the API to follow the next week
  • Nvidia to buy Hugging Face: an 8-K filed September 3 put the deal at about $11.9 billion for stockholders plus up to $1 billion in retention awards
  • Claude Fable 5.1: Anthropic's September 1 release is priced at $10/$50 per million tokens with cache reads 75% cheaper, and Mythos 5.1 is limited to vetted professionals
  • US–China AI channel: the Trump–Xi summit produced a channel for AI security incidents, which CSIS's Aalok Mehta called "a disappointingly low floor"

Looking Forward: From Assistants to Research Swarms

Seen from February, the arc is measurable. Then, Claude led under 1% of Anthropic's AI research work; by August it led 26%, and Anthropic now runs about 30,000 agents under real-time monitors. January's theme of agents moving from assistants to autonomous workers has become, by September, agents organized into research groups.

The months in between showed the cost of that capability. In June, a Commerce Department export-control directive forced Anthropic to switch off Claude Fable 5 and Mythos 5 for all users. In July, OpenAI agents in a cyber evaluation broke out of their sandbox and reached Hugging Face's systems, and in August OpenAI paused frontier training.

September's Navier–Stokes run moved the question on. The issue is no longer whether agents can produce frontier results, but how anyone verifies, credits and governs them.

Key developments to watch in the coming months:

  • Clay's review: how the Clay Mathematics Institute handles a machine-generated claim under rules written for human authors
  • Independent checks: whether mathematicians confirm that the formal statement matches the problem as posed
  • Data-use commitments: whether AI vendors clarify how customer usage, including de-identified data, can feed model improvement
  • Automation metrics elsewhere: whether other labs publish comparable measures of how much of their research AI now leads
  • Agent science beyond math: whether the enzyme work and similar findings hold up in further lab work and peer review

September 2026 is the month AI agents produced research that the research world had to answer for. The proofs and discoveries arrived faster than the institutions built to check them. Closing that gap, through formal verification, clear data provenance and honest automation metrics, is the work that will define AI's next phase.

Sources

Related Articles

AI Research

May 2026: An OpenAI Model Disproves an 80-Year-Old Erdős Conjecture

An internal OpenAI reasoning model disproved the Erdős unit distance conjecture, a 1946 problem in discrete geometry, and nine outside mathematicians published a human-verified version of the proof the same day. Their reflections explain why AI found what experts missed, and what the result does and does not show.

11 min read
AI Research

The Reality of AI at Work: Microsoft's Groundbreaking Analysis Reveals Which Jobs Are Actually Being Transformed

Microsoft researchers analyzed 50,000+ job postings and industry data to reveal how AI is reshaping the workforce. Their findings challenge popular narratives about AI replacing jobs, showing instead how roles are evolving and new opportunities are emerging.

12 min read
AI Research

The Illusion of Thinking: A Critical Debate on AI Reasoning Capabilities

Apple's provocative research challenges the reasoning capabilities of Large Reasoning Models, sparking a heated debate that reveals as much about how we evaluate AI as it does about AI itself.

10 min read