$2,000 of Compute Cracks Problems Mathematicians Left Open for 40 Years
In the early hours of August 2, OpenAI quietly released a 249-page paper, and the mathematics community has been arguing about it ever since. The paper is not a benchmark leaderboard and not a capability demo. It announces Astra, OpenAI's next-generation model, and credits it with resolving ten long-standing open problems in mathematics and theoretical computer science in one sweep.
The ten problems span eight fields: high-dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics, among others. Each has sat open for at least a decade, and some for over forty years — high-dimensional sphere packing had not moved since 1978. Astra produced these results on roughly $2,000 of compute.
Not an Assistant This Time — the Author
AI's role in mathematics has been shifting for two years: assisting symbolic computation, checking proof steps, walking students through calculus. All of that is tooling. What Astra did is structurally different:
- The AI generates the mathematics. Not retrieving known results, not stitching citations — constructing new proof paths and counterexamples.
- Humans organize the manuscript. The arguments arrive from the model; people shape them into standard academic form.
- The AI formalizes its own work. Astra translates its arguments into Lean and produces machine-checkable proof certificates.
OpenAI's signed statement is unusually direct: the mathematical arguments were generated entirely by the AI system, with the team assisting on the manuscript and formal verification and taking responsibility for correctness. In the sense that matters, Astra is the first author.
The Standout Results
All ten results are serious. Four deserve individual attention.
A proof that non-sofic groups exist
Gromov asked in 1999 whether every countable group is sofic — a question that has anchored geometric group theory for a generation. Astra answers no, and goes beyond a single counterexample: it constructs an infinite family of countable groups, pairwise distinct in structure, all non-sofic. A Caltech-trained mathematician called it Fields-medal-grade work, and Epoch AI's OpenMath evaluation rated it a breakthrough of the year.
The Connes rigidity conjecture falls
The Connes rigidity conjecture has been a central target in operator algebras since it was posed. Astra did not merely find one counterexample — it built a batch of groups that differ structurally while sharing identical von Neumann algebras. The picture one commenter used: someone claims no two snowflakes are alike, and the model answers with a blizzard — every flake distinct on the outside, every flake identical at the core.
A new upper bound for high-dimensional sphere packing
The first improvement since 1978. How densely spheres can be packed in high-dimensional space is not a puzzle for its own sake — the same mathematics bounds error-correcting codes and shapes the design of communication systems.
A superexponential lower bound for multicolor Ramsey numbers
This settles problem #183 from Erdős's list. Ramsey theory is bedrock combinatorics, and the result signals something larger: the model's abstract reasoning in pure mathematics now operates at the professional frontier.
The Method Is the Real Headline
The specific theorems matter less than the workflow Astra executed. The classical research loop runs hypothesis, experiment, analysis, paper, verification. Astra's loop looks like this: the model generates hypotheses and proof paths; humans evaluate, filter, and organize; the model completes formal verification on its own. AI owns the two most time-consuming stages — generation and verification — while people concentrate on judgment and integration.
Just as notable is the breadth. Astra was not tuned to one problem family; it landed blows across eight different mathematical directions. This is no longer "AI caught up to human mathematicians in one niche" — it is a system exhibiting cross-domain discovery.
OpenAI paired the release with a rollout: ChatGPT for Academic Researchers, free access to top-tier models for 100,000 scientists and mathematicians. That is not a coincidence. It is infrastructure going in ahead of a new research paradigm.
What $2,000 Actually Buys
The compute figure deserves a second look. A PhD student spending five years in one of these fields costs a university well over $200,000 once you price tuition support, advisor time, compute resources, and conference travel. Astra burned roughly $2,000 of electricity to traverse proof paths that would occupy a human mathematician for years.
Honesty requires the caveats. The results still face community peer review, and OpenAI has not disclosed how many attempts failed. If each success cost $2,000 and took a hundred tries, the real bill is $200,000 — the savings, not the sticker price, is the story. But the direction is unambiguous: AI is turning from an answer generator into a hypothesis generator. In fields that run dense search-and-verify loops — mathematics, theoretical computer science, drug discovery, materials science — that shift may arrive faster than most people expect.
What It Means for AI Practitioners
- The research-agent roadmap is now visible. Astra's core capability is not single-shot reasoning but long-horizon collaboration among multiple agents on one hard problem — the step that turns conversational tools into research partners.
- The cost curve is steepening. Ten frontier problems for $2,000 means AI-driven discovery is becoming mass-producible rather than a laboratory luxury. If your work involves literature review, hypothesis generation, or pattern detection, the integration question is already overdue.
- Formal verification is becoming the standard. OpenAI used Lean for machine checking, and that is no accident. As AI-generated arguments grow more complex, machine-verifiable output will be the baseline requirement for AI research. Lean fluency may be the most covetable skill of the next few years — the way Python was in the last decade.
- The timing is deliberate. Altman has demonstrated Astra at the White House, and the new US administration's AI framework deadline falls this weekend. As one of the first models under review, Astra's release cadence will shape how the entire industry reads AI regulation.
$2,000. 249 pages. Ten open problems, some untouched for forty years.
This is not the usual "models got better" news cycle. It is the first time an AI system has submitted work to the mathematical community as an independent author. Mathematician Thomas Bloom kept his praise measured but pointed: this breakthrough outweighs OpenAI's earlier disproof of the unit distance conjecture.
That earlier result landed in May 2026. Three months ago.
The speed is the most unsettling part of this story.