What OpenAI Navier-Stokes 10,000-Agent Run Really Cost
OpenAI says 10,000 agents solved Navier-Stokes in 88 hours on 130B tokens. The list-price math points to $6.5M in output alone.
OpenAI says roughly 10,000 coordinating agents resolved the forced Navier-Stokes existence and smoothness problem in about 88 hours, exchanging 2.7 million messages and generating approximately 130 billion output tokens. At public GPT-6 Astra output rates, that output alone lists near $6.5 million, with the full multi-problem effort listing near $15 million.
What OpenAI actually claimed
The claim lives in a September 8, 2026 OpenAI research post on the Navier-Stokes Millennium Prize problem, with a writeup plus a Lean formalization on GitHub. Treat it as an unpublished company claim, not an accepted result. The Clay Mathematics Institute, which lists Navier-Stokes among its Millennium Prize problems, has not recognized the proof, and OpenAI states it does not intend to claim the $1 million prize.
The run, as OpenAI describes it: training of a new internal model began August 28, 2026, a model the company says is significantly more capable than GPT-6 Astra. On September 1, after hearing rumors that two Millennium Prize problems had been resolved, OpenAI pointed groups of coordinating agents at every open Millennium problem plus some easier related problems. Agents could run code and read a cached version of the internet inside isolated environments, with communication inside each group.
An easier warmup paid off first. Nearly 100 agents spent about 50 hours producing a disproof of the regularity question for the unforced Euler equations (the viscosity term removed). OpenAI then shifted agents off the other Millennium problems, fed the Euler result back in as guidance, cross-pollinated groups by using Codex to consolidate insights across agent groups, and concentrated on Navier-Stokes. The peak group involved on the order of 10,000 concurrent agents. The agents reached the result on September 5, about 88 hours after launch. GPT-6 Astra then spent another 17 hours turning the argument into a Lean formalization and checking it.
The result itself: a proof that an initially smooth fluid at rest, with a smooth applied force and finite energy throughout, can develop a singularity (velocities growing without bound) in finite time. That establishes statements C and D in the official Fefferman formulation, the disproof direction. Two of the four statements, in other words, resolved in the negative.
| Item | Value |
|---|---|
| Navier-Stokes output | 130B |
| All attempted problems | 300B |
The token math, line by line
Start with the two hard numbers from the OpenAI post: approximately 130 billion output tokens and 2.7 million messages for Navier-Stokes, versus about 300 billion output tokens and 4.9 million messages across all attempted problems. Everything below reprices those counts at public rates. That repricing is a proxy, not the true cost, because the run used an unreleased internal model, and internal inference cost is not public.
The reference rate comes from OpenAI developer documentation for GPT-6 Astra, the closest public model: $10 per million input tokens and $50 per million output tokens at Standard short-context rates, with higher long-context rates above 272,000 input tokens. Multiply through:
- Navier-Stokes output: 130,000 million-token units at $50 lists at $6.5 million.
- Full effort output: 300,000 units at $50 lists at $15 million.
Those two products explain the two headlines circulating since September 9. The BBC derived roughly $10 million from OpenAI output pricing for the Navier-Stokes run, a figure consistent with output plus a material input bill. New Scientist headlined the full effort as $15 million of AI effort, which matches the all-problems output at list price before inputs. OpenAI research head Mark Chen put the cost only as “in the millions,” as reported by AFP.
Three caveats keep the math honest. First, input tokens are unbilled in every public telling: 2.7 million Navier-Stokes messages each carried prompts, tool results, and cross-pollinated Codex summaries, and none of that input volume is disclosed. At $10 per million, even modest per-message inputs add millions more. Second, long-context pricing bites exactly this kind of workload: agent handoffs with big shared contexts can cross the 272K threshold where output bills at 1.5x across the whole request, which would push the proxy up, not down. Third, internal cost is almost certainly below list price (no retail margin, possible batch or off-peak scheduling), while the failed-branch spend is real either way: roughly 170 billion tokens went to Millennium problems that did not resolve, and that spend bought information (Navier-Stokes looked promising) rather than output.
| Step | Change | Running total |
|---|---|---|
| All-problems output (300B tokens at $50/M) | $15M | $15M |
| Other problems (170B tokens, no resolution) | -$8.5M | $15M |
| Navier-Stokes output (130B tokens) | $6.5M | $6.5M |
| Inputs and verification (illustrative, undisclosed) | +$1.5M | $8M |
| List-price proxy (near BBC ~$10M derivation) | $8M | $8M |
Why the token bill is only the first line
Four cost layers sit outside the token count, and each generalizes to anyone pricing agent swarms.
First, failed branches are the strategy, not waste. OpenAI tried every open Millennium problem in parallel, then killed all but the most promising. Anyone running agent fleets should budget the same way: the winning branch cost $6.5 million in output, but finding it cost $15 million. Portfolio math, not single-shot math. The build versus buy harness analysis makes the same point at startup scale: orchestration overhead decides the bill before model choice does.
Second, coordination has its own meter. Codex consolidation passes, cached-internet reads, code-execution sandboxes, and 2.7 million inter-agent messages each consume inputs that never appear in an output-token headline. This is the harness tax documented in harness engineering techniques for 2026: cross-group pollination improves results and multiplies spend simultaneously.
Third, verification is a second bill. Seventeen hours of GPT-6 Astra formalizing and checking the proof in Lean is compute, plus the human mathematicians who supervised, plus the isolated-evaluation safeguards OpenAI says ran throughout. Formal verification is the expensive kind of confidence, and it scales with proof size, not with how fast the answer arrived.
Fourth, calendar compression costs money even when tokens are fixed. Ten thousand concurrent agents finishing in 88 hours is a latency play: reserved capacity, peak concurrency, and the Euler pilot (100 agents for 50 hours) that de-risked the main run. Faster answers cost more than the same tokens stretched over a month. Anyone modeling monthly agent spend on steady-state rates will underprice burst work by an order of magnitude.
The credit dispute, stated neutrally
NYU professor Tristan Buckmaster says he and Anthropic researcher Levent Alpoge, in a personal collaboration, had been pursuing the same forced-blowup route for months, logging their work in OpenAI Codex, and that OpenAI accelerated after learning of their progress. His personal statement PDF stresses what he is not claiming: he writes that he has not seen the OpenAI proof, does not know what its model did, and is not accusing anyone of anything, while asking whether Codex session data was accessed or trained on.
OpenAI denies accessing their work, says neither researchers nor agents saw it before public release, and published a September 10 update saying its investigation found Buckmaster Codex prompts could not have influenced the system, including through training. It adds the caveat that it cannot rule out de-identified usage data having improved its models generally, recognizes the priority of the Alpoge and Buckmaster work on forced Euler, and notes the proofs differ (theirs: forced Euler; the OpenAI system: unforced Euler plus forced Navier-Stokes). WIRED reporting and CNBC reporting each carry both sides. The Capital and Compute rule here: every compute figure above traces to OpenAI or to arithmetic on public rates, and the priority question stays open until independent mathematicians verify both proofs.
What builders should take away
Price the portfolio, not the winner. The per-token trend (tracked here) keeps falling, yet frontier bursts keep getting bigger: 130 billion output tokens in under four days. Unit costs drop while flagship attempts concentrate spend. Budget both curves.
Treat verification as a line item from day one. A Lean-checked proof is the gold standard, and it added 17 hours of the strongest available model plus human oversight. For ordinary agent deployments the analogue is evals, graders, and human review: the question of whether GPT-6 Astra counts as AGI matters less than whether each agent output gets checked by something stronger than the agent that produced it.
Finally, log provenance before it matters. The dispute above is a preview of every future argument about whose prompts trained whose model. Teams running evals on third-party harnesses should know where their traces go, what the provider retains, and what “de-identified” covers, before the run that matters.
Bottom line
A 10,000-agent, 88-hour math run lists near $6.5 million in output tokens alone at public rates, near $8 to $10 million with inputs and verification, inside a $15 million all-problems portfolio. That is the price of compressing months of mathematical search into four days with massive parallelism, a failed-branch portfolio, and machine-checked verification. Whether the proof survives scrutiny is a separate question for mathematicians. The cost structure is already legible, and it reprices what burst-scale agent work goes for.
Sources
- OpenAI (2026). On the Navier-Stokes Millennium Prize Problem. Research post, September 8. https://openai.com/index/navier-stokes-solution/
- OpenAI (2026). GPT-6 Astra developer documentation: pricing. Checked September 2026. https://developers.openai.com/api/docs/models/gpt-6-astra
- Clay Mathematics Institute. Navier-Stokes Equation (Millennium Prize problem page). https://www.claymath.org/millennium/navier-stokes-equation/
- Charles L. Fefferman. Existence and Smoothness of the Navier-Stokes Equation (official problem description). Clay Mathematics Institute. https://www.claymath.org/wp-content/uploads/2022/06/navierstokes.pdf
- Tristan Buckmaster (2026). Personal statement on concurrent work (PDF). NYU. https://cims.nyu.edu/~tristanb/statement.pdf
- Matthew Sparkes (2026). OpenAI has solved the Navier-Stokes Millennium problem using $15m of AI effort. New Scientist, September 8. https://www.newscientist.com/article/2588063-openai-has-solved-the-navier-stokes-millennium-problem-using-15m-of-ai-effort/
- Kali Hays (2026). OpenAI says it cracked 90-year-old maths problem in 88 hours. BBC News, September 8. https://www.bbc.co.uk/news/articles/cy7zygy3rl2o
- Ian Sample and Dan Milmo (2026). OpenAI claims to have solved maths problem that stumped humans for decades. The Guardian, September 8. https://www.theguardian.com/science/2026/sep/08/openai-claims-to-have-solved-maths-problem-that-stumped-humans-for-decades
- Ville Satama (2026). OpenAI Just Claimed a Huge Math Discovery. Some Academics Are Crying Foul. WIRED. https://www.wired.com/story/openai-navier-stokes-math-discovery-academics
- Katrina Bishop (2026). OpenAI claims to have solved the 90-year-old Navier-Stokes math problem in 88 hours. CNBC, September 9. https://www.cnbc.com/2026/09/09/openai-navier-stokes-math-problem-solved.html
Frequently asked questions
- How much did the OpenAI Navier-Stokes run cost?
- OpenAI only says millions. Repriced at public GPT-6 Astra output rates ($50 per million), the reported 130 billion Navier-Stokes output tokens list near $6.5 million, and the 300 billion tokens across all attempted problems list near $15 million. Inputs, long-context uplifts, and verification sit on top, which is consistent with the BBC derived figure near $10 million.
- How many tokens and messages did the run use?
- For Navier-Stokes: approximately 130 billion output tokens across 2.7 million agent messages, reaching the result about 88 hours after launch with up to 10,000 concurrent agents. Across all attempted Millennium problems: about 300 billion output tokens and 4.9 million messages. GPT-6 Astra then spent 17 more hours on Lean formalization.
- Is the Navier-Stokes proof accepted?
- No. It is an unpublished company claim with a released writeup and Lean formalization, not yet independently verified or recognized by the Clay Mathematics Institute. OpenAI states it does not intend to claim the $1 million prize.
- What is the dispute with Buckmaster and Alpoge about?
- Tristan Buckmaster (NYU) and Levent Alpoge (Anthropic) published related machine-checkable results hours earlier and questioned whether OpenAI accessed their Codex session data. OpenAI denies accessing their work and published investigation findings to that effect, while saying it cannot rule out that de-identified usage data improved its models generally.