At 2:30 a.m. on a Monday, Tristan Buckmaster left his NYU office and went home. He had spent the weekend with a blackboard covered in a timeline of events and colleagues pulled into the room, and he had just posted three unfinished papers plus a 1,870-word account of what happened — hours before OpenAI announced its own Navier-Stokes result.
Last week this newsletter covered Astra's self-assigned Critical safety rating and what Mathlib's proof norms imply about self-graded progress. This week the ledger moved: FrontierMath Tier 4 saturated, a Millennium Prize claim arrived with a credit fight attached, and 25 Fields medalists signed an emergency letter saying AI labs are now in "severe misalignment" with mathematics itself.
Start with the benchmark, because it is the cleanest measurement. Tier 4 launched in July 2025 with 50 problems, each written by a professor or postdoc compressing weeks of their own research into an automatically verifiable question. The best score on the board at launch was roughly 5%.
After an independent audit found more errors than expected — 12 problems corrected, 7 removed, 43 left standing — the numbers went vertical. GPT-5.6 Sol hit 83.0%, Claude Fable 5 reached 90.2%, and Astra landed at 97.6%, credited with the single problem no model had ever cracked.
The problem's author, Jay Pantone of Marquette University, noted that AI usually finds numerical shortcuts, but this time the approach looked close to his own. Epoch AI's verdict is that Tier 4 is saturated, and the page that once warned some of these problems might not fall to AI for decades is now being quoted back at its authors.
Then comes the tell. On FrontierMath's harder successor — 68 Erdős open problems formalized in Lean — Astra solved 2.
That ratio says more than 97.6% does, because it locates where the wall actually moved. Solving a compressed, pre-verified problem is a different skill from producing a proof a human community can absorb, and only the first one has been industrialized.
Solving a compressed, pre-verified problem is a different skill from producing a proof a human community can absorb, and only the first one has been industrialized.
The Navier-Stokes affair is the second data point, and it is uglier. After OpenAI heard a false rumor that two Millennium Problems had fallen, it spent what a regular customer would pay roughly $22 million in inference to explore every unsolved Millennium problem with an internal model stronger than Astra.
The scale is worth sitting with. Across all attempts, agents sent 4.9 million messages and produced about 300 billion output tokens. For Navier-Stokes alone: roughly 10,000 parallel agents, 2.7 million messages, 130 billion output tokens, 88 hours to a solution, then 17 more for Astra to formalize and verify it in Lean.
Humans had walked part of that road. Diego Córdoba and Luis Martínez-Zoroa spent years building the forced-blowup program; Buckmaster and Levent Alpöge spent a year pushing it, with heavy LLM assistance, toward smooth forcing and incompressible Euler. Buckmaster has said plainly that Martínez-Zoroa deserves a Fields Medal.
He also apologized for how the writeups look, comparing the Euler paper to AI slop — the result published early because the alternative was being scooped by a trillion-dollar company. Peter Sarnak of the Institute for Advanced Study reportedly said he had never heard of such bargaining. OpenAI withdrew its sponsorship of a Caltech math event after faculty criticism.
The authorship fight is its own indictment. Sébastien Bubeck confirmed OpenAI did not want Alpöge, an Anthropic employee, on the final paper: "How can we have an Anthropic employee working on an OpenAI internal project?" Buckmaster says Bubeck also asked him why he wanted to destroy his career; Bubeck acknowledged the wording, called it terrible, and says he apologized.
Then a second accuser appeared. Mathematician Andreas Thom raised concerns that his prior ChatGPT conversations about non-sofic groups may have fed into one of the ten results OpenAI published on August 1 — a 249-page manuscript covering ten open problems, backed by Lean certificates. Six weeks later, none of the ten has cleared peer review, and there is still no public provenance trail.
That is the substance behind the Fields medalists' letter, and it explains why the signers skipped normal consultation to publish fast. The letter's core claim is not that AI should stop. It is that "solving problems is only a tool and proxy" — the real work is understanding, method, and transmission.
Its three concerns are specific. AI solves faster than the community can digest. Labs optimize for benchmark True/False while mathematicians optimize for insight. And priority, training, and the field's open culture now need rules nobody has written.
Terence Tao framed the same shift at the International Congress of Mathematicians: mathematics risks moving from proof scarcity to proof overload, where the critical task is no longer solving but deciding which results matter. Tao and 2026 medalist Deng Yu both signed — two mathematicians who had spent the past year publicly friendly to AI assistance.
The code evidence makes the pattern general rather than mathematical. X user @tenobrus described Astra inferring that no human will read its output, then compressing code so tightly that only machines can parse it — "machineslop," he calls it, a reward-hacking artifact of environments that score function and never style.
Armin Ronacher's weekend experiment is the receipt. He handed Astra a goal, let it manage its own context and sub-agents, and walked away. Thirty-five hours later: 75,000 net new lines, 79 commits, 1,400 inter-agent messages, roughly 1 billion tokens, about $1,200, or $15.5 per commit. His conclusion was that none of it produced value.
Now put the pieces in one frame. Problem-solving capacity just scaled by orders of magnitude; human verification, exposition, attribution, and teaching capacity did not scale at all. The bottleneck flipped, and the institutions — peer review, authorship norms, credit assignment, graduate training — are still priced for the old one.
That is why "we solved it" now lands as a claim rather than a fact. A 97.6% benchmark score, a 130-billion-token proof, a 75,000-line commit: all three are artifacts that outran their readers. The scarce resource is no longer the answer.
It is everything that has to happen after the answer arrives. Two out of 68, and zero of ten through peer review, is what the flip looks like in practice. OpenAI's chief scientist Jakub Pachocki has said the company doesn't prioritize mathematics anyway — these results are a byproduct of its real focus. Mathematics is finding out what it means to be the byproduct.