A 27-year-old conjecture about binary matrices was cracked by ChatGPT 5.5 Pro and 100,000 lines of Lean code. The solver was not a human.
The three-punch proof
In the first half of 2026, AI systems solved open mathematical problems at a rate that rewrites the funding formula for pure research. Three events landed in sequence, each from a different lab, each using a different architecture, each producing a verifiable result.
On February 5, Google DeepMind released the FirstProof challenge—ten problems drawn from working mathematics—and gave entrants eight days. Aletheia, DeepMind's math agent powered by Gemini Deep Think, solved six of them. The assessments came from human experts; the six were judged rigorous proofs.
In the same window, an independent system called QED—a multi-agent pipeline orchestrating Claude, Codex, and Gemini through their command-line interfaces—cracked an open problem in partial differential equations. Three separate institutions verified the result.
Then in June, a researcher using ChatGPT 5.5 Pro for proof generation and Codex for verification resolved the KTV conjecture, a 27-year-old open problem on the mixing properties of the swap chain over binary matrices. The Lean formalization alone ran for roughly 100 hours and produced an initial 100,000 lines of code.
Three labs. Three problems. No coordinated effort. The direction of travel is set.
The lemmas and the logbook
FirstProof problems are lemmas—intermediate technical statements that working mathematicians need, not standalone conjectures with names like "Riemann." At least one, Problem 7, had been previously described as an open problem of interest. For Aletheia to hit six of ten in a week is the equivalent of an AI system doing a productive postdoc's quarter of work while the postdoc sleeps.
The PDE problem QED solved was not a toy. The system takes a LaTeX problem statement and produces a rigorous natural-language proof, routing sub-tasks across models. The human in the loop reviewed the final proof. Three institution-level verifiers signed off.
The KTV conjecture is deeper still. It asks whether the swap chain—a Markov chain used to sample random binary matrices with fixed row and column sums—mixes rapidly for every feasible choice of margins. The AI-generated proof delivered a worst-case optimal spectral-gap bound. The human researcher who ran the experiment described what the machine produced: "The proof itself took only a handful of prompting rounds. The Lean formalization took Codex about 100 hours, consumed an entire week of my Pro quota, and initially produced roughly 100,000 lines of code."
A week of Pro quota is not free. It is also not even close to the cost of a graduate stipend, let alone a postdoc salary.
The auto-regressive mathemagician
The new math machinery works through a common loop: an LLM generates candidate proofs or proof steps, a formal verification engine—usually Lean—checks them, errors are fed back into the model, and the cycle repeats. Aletheia iterates in natural language. QED splits the work across multiple models. The KTV solver leaned on GPT-5.5 Pro to generate the proof and Codex to hammer it into Lean.
The human role collapsed from "mathematician" to "prompt engineer who understands the problem well enough to know which output to believe." The validation bottleneck is now software. Lean either accepts the proof or it does not. Human intuition, the secret weapon of mathematics for millennia, is not part of the loop.
The cost asymmetry is the kind that reshapes institutions. Producing 100,000 lines of verified Lean code for the price of a single Pro subscription week is roughly two orders of magnitude cheaper than supporting a postdoc for the months it would take to attempt the same problem—with no guarantee of success. The KTV solver noted that the formalization consumed a full week of quota. That week cost less than dinner for two at a decent restaurant.
The theorem market inverts
The institutions that matter in mathematical discovery are about to change. Look at who built Aletheia, who trained the models QED orchestrates, whose inference capacity underlies the KTV proof. Google DeepMind. OpenAI. Anthropic. These are not mathematics departments. They are compute organizations.
The 2026 results demonstrate a brutal new unit-economics: a mid-level research lemma costs less to generate via AI pipeline than to assign to a human. The Aletheia FirstProof run, the QED PDE proof, the KTV resolution—each landed at a cost measured in subscription fees and compute cycles. That is the first domino.
The second is the trajectory. It is log-linear. The cost per lemma will halve, then halve again, as models improve and orchestration tightens. Within 12 to 18 months, pipelines will tackle problems that today feel genuinely deep: unsolved conjectures in combinatorics, PDE theory, algebra. Millennium problems are a different category requiring conceptual leaps; the ROI there remains poor. But the set of problems that yield to a search-verify loop is larger than most mathematicians believe.
This flips the competitive dynamic. The institution with the most capable reasoning models and the largest inference budget—Google DeepMind, OpenAI, Anthropic—sits on the discovery frontier. A mathematics department that spends its budget on postdocs rather than compute will be out-produced, lemma for lemma, within two years. Research funding will reallocate from pure human talent to compute budgets, not because deans want it to, but because the output-per-dollar gap will become too wide to ignore.
Peer review breaks next. No human referee can audit a 100,000-line Lean proof. Journals will face a choice: accept machine-verified proofs with AI co-authors, or watch the cutting edge publish on arXiv and never look back. Within 12 months, at least one peer-reviewed journal will accept a paper with an AI listed as co-author. The formal verification is the review; the human referee's role reduces to checking that the theorem statement is correctly translated into the formal language.
Then the cost floor drops out. By mid-2027, generating a new proof will cost less than hiring a postdoc. Funding agencies designed for an era of expensive human labor—NSF, EPSRC, ERC—will get outbid by tech companies who see mathematical output as a byproduct of inference capacity they already own. The theorem market inverts. The buyer is no longer the grant committee. The buyer is the lab with the biggest inference budget.
The system that tracks this shift with uncomfortable clarity is already public. Andrej Karpathy's autoresearch repository—a project where an AI agent autonomously experiments with LLM training code overnight—has accumulated 93,394 stars, 13,270 forks, and 196 open issues. Three files. Autonomous research as a weekend project. The frontier is not hiding in a lab. It is on GitHub, and thousands of researchers are forking it.
Compute over chalk
Mathematics department chairs, PhD students, and journal editors all face the same question with different answers.
For the department chair: If your curriculum does not integrate Lean or an equivalent formal verification tool within the next 18 months, your graduates will compete with pipelines that cost less than their stipends and produce verified output faster than they can write a first draft.
For the PhD student: Learn to operate the loop. The mathematician who can formulate problems for an AI pipeline and interpret the output will produce more than the one who cannot. This is not a threat to mathematical creativity. It is a shift in where creativity applies—higher up the abstraction stack, away from the intermediate lemmas the machines now own.
For the journal editor: Start building the editorial workflow for AI-co-authored, machine-verified papers. The first journal to get this right captures the pipeline of results that will otherwise never pass through a traditional review process.
For everyone else: Bet on compute, not human hours. The AI-as-calculator era ended the week Aletheia hit six of ten on FirstProof. The new era treats formal reasoning as a compute problem. Organizations that buy inference capacity will own the discovery front. Organizations that buy chalk will read about the results.
The temperature check
The binary matrix verdict landed in June 2026, delivered by a system that did not pause, did not consult a colleague, and did not care that the problem had stood for twenty-seven years. It generated a proof, formalized it in Lean, and moved on to the next token.
One researcher, one Pro subscription, one hundred hours of Codex time. The era of the lonely genius in a garret ended not with a bang but with a pull request. The era that follows belongs to the lab with the largest inference budget.
