OpenAI’s Astra: Ten math breakthroughs for about $2,000

Ten claimed results, Lean certificates and a $2,000 inference bill point to a new bottleneck: expert attention, not proof generation.

$OpenAI’s Astra: Ten math breakthroughs for about $2,000$

I’m staring at a 249-page mathematical manuscript whose discoveries allegedly cost about $2,000 in model inference. OpenAI’s Astra claims ten breakthroughs on long-unsolved math problems, for less than some startups spend on a two-day offsite with mediocre focaccia and one man named Chad explaining alignment.

Before anyone orders a commemorative Fields Medal, these are OpenAI’s claims, published on August 1, 2026. A company blog, a giant PDF and a GitHub repository full of Lean certificates do not create instant mathematical consensus. Specialists still have to inspect whether each formal statement captures the historic problem and whether the informal interpretation survives contact with other mathematicians.

Even with that caveat, the economics are nuts. Serious AI-generated proofs may now cost a few thousand dollars to produce and can be checked by software. Qualified human attention is suddenly the expensive ingredient.

I’ve spent 20 years shipping technology through Ad Astrum, including home-automation systems for E.ON and Megafon, plus a connected Pascucci espresso machine that sent brew telemetry to the cloud. Yes, we put coffee on the internet. Italy has forgiven me, although my mother remains undecided.

Every time production gets dramatically cheaper, volume goes feral. Teams produce more updates, more features and far more junk nobody requested. Mathematics is about to learn the same lesson, only with conjectures instead of push notifications.

Ten results look like a production run

One flashy theorem can be a demo engineered for applause. Ten results across ten specialties look like the first shift at a factory.

OpenAI described its August 1 batch this way:

Today, we are sharing a selection of ten results, each of which resolves or makes substantial progress on a long-standing open problem.

The breadth made me sit up. Astra allegedly constructed a non-sofic group, answering a central question in group theory, and produced a counterexample to Connes’s rigidity conjecture in operator algebras.

I cannot referee operator algebras over an aperitivo. Almost nobody can referee all ten areas, which makes this batch an unusually messy review job.

The list includes a superexponential lower bound for multicolor triangle Ramsey numbers, resolving Erdős problem 183. Astra also allegedly resolved Erdős problems 146 and 180 through counterexamples involving compactness and degeneracy conjectures in extremal graph theory.

Then we get arithmetic circuit complexity. OpenAI claims an arithmetic-formula lower bound of order $n^4/\log n$ for computing the permanent, documented in a Lean module called Permanent.lean. That sentence will make a complexity theorist lean toward the screen. A normal person will check whether lunch has arrived. Both reactions are valid.

Other claimed advances cover high-dimensional sphere packing, binary and spherical codes, quantum games, lattice cryptography and convex geometry. The quantum result is especially broad: exponential parallel repetition for arbitrary finite two-player quantum games.

These are different species of mathematical work. Some kill conjectures with counterexamples. Others improve quantitative bounds or establish structural theorems. Ehrhart’s volume conjecture gets an exact extremal answer in every dimension.

OpenAI’s previous major reveal came in May 2026, when an unreleased model disproved the Erdős unit-distance conjecture, posed in 1946. According to OpenAI’s August publication, that result has already inspired subsequent work by Thomas Bloom, Will Sawin, Andrew V. Sutherland Schildkraut, Dmitrii Zhelezov and Cosmin Pohoata, among others.

In a July 2026 article republished by Phys.org, Trefor Bazett wrote that human researchers adapted the central technique within a week to attack the sum-product conjecture. That is what useful discovery looks like after the press release: mathematicians grab the weird new tool and start hitting nearby problems with it.

The $2,000 figure changes the queue

OpenAI says Astra generated the mathematical arguments first. Humans then prepared the manuscripts using the same model. Astra subsequently formalized each argument into a Lean certificate, and OpenAI released reasoning walkthroughs that describe how the ideas allegedly came together.

Its cost claim is unusually specific:

The results were achieved by an internal version of Astra, our next major model. The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.

That figure deserves a red asterisk large enough to see from Sicily. It excludes Astra’s training, researcher salaries, infrastructure costs, failed internal experiments and whatever heroic debugging happened at 3:17 a.m. The $2,000 covers marginal inference at Sol API rates. Building the factory cost considerably more.

Once the factory exists, marginal cost controls volume.

I learned this in the least glamorous possible way while building connected products. With ALYT, a US home-automation platform, every reduction in firmware-update or cloud-deployment costs encouraged us to ship more often. App-store rules and device compatibility then clogged the process. Field testing did the rest. Cheap deployment moved the queue somewhere else.

Software has already been through this cycle. Open-source packages cut the cost of assembling applications. AWS turned servers into metered infrastructure. GitHub Copilot and Codex now produce code faster than many teams can review sensibly, a situation familiar to anyone who has merged a pull request on Friday and discovered religion on Saturday.

Mathematics has an effectively infinite backlog. Paul Erdős alone left hundreds of open problems. Add decades of conjectures from thousands of researchers, unpublished questions sitting in notebooks and every new variation Astra can formulate after breakfast.

The model’s operating environment matters too. In OpenAI’s July 29 report on ARC-AGI-3, GPT-5.6 Sol scored 13.3 percent using the official harness. Retained reasoning and context compaction raised the score to 38.3 percent while cutting output tokens by a factor of six.

OpenAI’s improved setup used a 175,000-token context limit. The model remembered earlier reasoning and preserved useful observations through compaction. It stopped rebuilding its mental map after every action. Very relatable. I also perform worse when somebody erases my whiteboard every 30 seconds.

The useful mathematician here is a complete system: Astra, persistent memory, managed context and formal verification, with search running through the whole thing. Judging that system from one naked chatbot prompt is like judging an IoT platform from one microservice before 100,000 devices wake up after a power outage.

OpenAI has announced free access to advanced ChatGPT models for 100,000 scientists and mathematicians. If researchers receive enough capability and usage, that becomes an enormous experiment in problem selection.

At $2,000 per generated proof and perhaps six months of expert attention to absorb it, mathematicians inherit the queue.

Astra's math breakthroughs showcased with graphs and equations, highlighting innovative tech solutions for affordable learning.

Suggested alt text: OpenAI Astra workflow showing an open problem moving through AI search, an informal argument, human manuscript preparation, a Lean certificate and expert interpretation, with approximately $2,000 in inference cost under the generation stage.

Lean checks the proof it receives

Lean is a proof assistant. A researcher encodes definitions and assumptions, then specifies a formal conclusion. Lean checks whether that conclusion follows under its logical rules.

OpenAI’s public ten-proofs repository contains separate files for all ten claims. They include NonSoficGroup.lean, ConnesRigidity.lean, Permanent.lean, GapCVP.lean, EhrhartVolumeInequality.lean and MulticolorTriangleRamsey.lean.

The repository uses Lean 4.32.0, mathlib and Lake. OpenAI provides commands for building the whole project or checking an individual module. This is considerably more useful than uploading a PDF and asking everyone to trust the vibes.

OpenAI described the workflow directly:

These arguments were then prepared into manuscripts by humans with the same model. Afterward, the model formalized each argument in a Lean certificate.

A successful Lean build establishes that the encoded conclusion follows from the encoded definitions and assumptions. Lean cannot independently decide whether those definitions match the object mathematicians meant when they posed the original conjecture.

Kevin Buzzard, a Lean maintainer and professor at Imperial College London, has been hammering on this distinction. In a July 20 post on the Xena Project blog, he explained how a formal proof can compile while targeting the wrong statement or encoding a definition that fails to capture the conventional mathematical object.

Reviewers therefore have separate jobs. They need to confirm the code compiles, inspect the formal statement and compare it with the historic conjecture. Then comes the question that shapes future research: what did the method contribute to the field?

A mathematically correct result can be narrow or hard to reuse. It can also be boring. Lean checks logic; it has no taste.

The scale is already silly. According to Buzzard, OpenAI’s Sol generated 1.2 million lines of Lean in three weeks while formalizing the earlier unit-distance work. Mathlib contained roughly 2.3 million lines built over nine years.

Three weeks versus nine years.

Buzzard and his postdoc Thomas Browning had previously inspected a formalization of the unit-distance argument produced by Logical Intelligence. That company was co-founded by Yann LeCun, while Fields Medalist Mike Freedman is its chief science officer. The field is filling up quickly. Logical Intelligence, Harmonic, Axiom AI, Logos Research and the frontier labs are all pushing on formal mathematics.

Buzzard also described receiving a 1,076-line Lean file for a counterexample involving finite free group schemes. Once he trusted the definitions and statement, he checked the file in under five minutes.

Five minutes sounds magical until the files arrive at industrial volume. At 400 files, quick semantic inspection becomes a full-time job. At 40,000, we have invented mathematical content moderation.

There is a security wrinkle. Lean is a programming language capable of running arbitrary commands. Buzzard ran generated code in a sandbox before trusting it. “Machine-checked” should never become the magic spell that makes smart people forget basic operational security.

Lean gives OpenAI’s ten Astra claims much more credibility than a model transcript would. It also allows formal correctness to scale faster than human confidence in what the formal object actually says.

Astra has factory-floor energy

Pop culture taught us to picture mathematical discovery as a cinematic flash. A tortured genius stares at a board, one emotionally significant piano chord plays, and somebody circles the answer.

Astra’s alleged output feels industrial.

Its apparent strengths include persistent search and recombination across specialties. It can also switch proof styles. A non-sofic group requires constructing an object. Connes’s conjecture falls through a counterexample. Sphere packing receives better bounds. Quantum parallel repetition gets a broad theorem.

Astra can pursue ugly branches without boredom, embarrassment or the urge to spend 40 minutes choosing the correct espresso bar. That helps.

The May unit-distance result offers the clearest example of recombination. Buzzard says its core connected a geometric conjecture from 1946 to the Golod–Shafarevich theorem from the 1960s, a deep result from number theory. Humans had possessed both ingredients for decades. The model allegedly found a bridge across specialist boundaries.

I have watched strong startup systems beat charismatic improvisation for 20 years. Founders love talking about vision because retry queues and telemetry look terrible in a keynote. The boring machinery keeps the company alive when the demo meets customers.

Astra’s advantage may come from industrialized persistence. It can scan more candidate connections, keep working through ugly branches and formalize the path that survives. Inspiration remains a lovely word. Throughput now has receipts.

The recent Jacobian-conjecture debate shows why the kind of result matters. Ott-Heinrich Keller formulated the conjecture in 1939, and Stephen Smale included it on his 1998 list of 18 major mathematical problems. In July 2026, Anthropic employee and Harvard mathematician Levent Alpoge announced an AI-assisted counterexample generated with Anthropic’s Fable model.

Andrew Blumberg told Mashable, as summarized by The Week, that a single counterexample can reveal essentially nothing about the surrounding theory. Fair. Finding one polynomial that breaks a universal claim differs intellectually from building an explanatory framework.

Melissa Lee’s analysis, also cited by The Week, points toward the machine advantage. An enormous space of polynomial mappings is perfect territory for computational search. One small, ugly object can still demolish 87 years of intuition.

I expect humans to be outcounterexampled before they are out-theorized. Machines can search huge spaces for one pathological object. Theory demands compression: explain why the object exists, find the mechanism that generalizes and decide which question should replace the dead conjecture.

That still wrecks the old workflow. Counterexamples steer entire fields. A machine can invalidate years of planned research before lunch without producing the next grand theory.

Mathematical prestige moves downstream

Terence Tao has described frontier AI as artificial general cleverness: broad, stochastic problem-solving that often relies on ad hoc methods. On his own AI views page, his recurring operational description is unreliable but powerful.

I like that framing. It skips the metaphysical food fight over whether a model genuinely understands mathematics. If the output kills an Erdős conjecture and survives formal checking, the department seminar has to deal with it.

Tao’s July 24 ICM 2026 lecture, Mathematics in the age of AI, focused heavily on the widening gap between proof generation and human comprehension. A verified proof can remain culturally orphaned when nobody understands the mechanism well enough to teach it or connect it to neighboring work. Somebody also has to formulate the next useful conjecture.

I’ll admit where I was wrong. I thought formal verification would mostly solve the AI-math credibility problem. Put the proof into Lean, compile it and move on.

Nope.

Formal verification solves a crucial logical problem. It also makes the shortage of interpretation impossible to ignore.

Jacob Tsimerman makes the labor impact harder to dismiss. The 2026 Fields Medal winner announced that he would take leave from the University of Toronto to work at OpenAI. The Atlantic compared the hire to putting Lionel Messi in a project-management role, which feels unfair to Tsimerman because I assume he has fewer opinions about Jira.

In his July interview with The Atlantic, Tsimerman forecast that AI could accelerate the production of interesting mathematics by factors of 10 or 100. He also warned that skills young Ph.D. researchers spend years acquiring may become less relevant.

That is brutal if you are 24, halfway through a doctorate and living on a stipend that makes Trader Joe’s frozen pasta feel aspirational.

Prestige will shift toward mathematical taste and synthesis. Journals may receive thousands of technically valid machine-generated results and face a question tougher than correctness: which ten deserve the community’s attention?

I’d bet on a future star mathematician becoming famous for choosing an important question, extracting reusable ideas from 400 machine proofs and making those ideas understandable. The theorem still matters. Placing it inside a living body of mathematics becomes scarce work.

Authorship gets weird quickly. OpenAI says Astra generated the arguments, while humans prepared the manuscripts and helped formalize the results.

The company’s stated position is unusually direct:

We helped prepare the manuscripts and formalize the proofs in Lean, and we take responsibility for their correctness, while the mathematical arguments themselves were generated by our system.

The Leiden Declaration on AI and Mathematics, signed by researchers from more than a dozen universities, has called for guardrails around attribution, transparency and peer review. Those are practical demands. A field built around intellectual lineage needs language for results generated by private systems and edited by humans.

Future researchers also need access to the discovery process. Attribution without reproducibility is branding with footnotes.

Cheap proofs still require an expensive factory

Astra remains internal and unreleased. Outsiders can inspect the selected manuscripts, walkthroughs and Lean certificates. They cannot rerun the full discovery process on their own questions.

We cannot see the failed attempts either. Nobody outside OpenAI knows how many problems Astra tried, how the ten were selected or which promising results remain private. The $2,000 estimate covers tokens used for successful solution generation. It reveals nothing about the cost of discovering which prompts and workflows yield publishable mathematics.

OpenAI consequently controls the question queue. It can influence which conjectures receive compute, which mathematicians get early access and which results become international news.

I have seen gentler versions of this dynamic in tech. Apple shaped mobile businesses through App Store ranking and review. Google shaped the web by deciding which pages people discovered. A lab operating the strongest private mathematician can shape science through problem selection.

OpenAI is already building institutions around that capability. Its ChatGPT for Academic Researchers initiative promises free access for 100,000 scientists and mathematicians, though advanced public models offer a different experience from Astra’s internal research setup.

Through the US Department of Energy’s Genesis Mission, OpenAI has committed $4 million in Codex access for approximately 2,000 researchers. It has also promised $3 million in API support for two focused scientific campaigns.

Genesis connects industry and universities with the Department of Energy’s 17 National Laboratories. OpenAI says more than 1,000 scientists across nine laboratories joined an AI Jam Session and tested frontier models on domain-specific problems. The company has also deployed reasoning models on Venado, the supercomputer at Los Alamos National Laboratory.

That is scientific infrastructure with access tiers and compute budgets. It also has national-strategy consequences. OpenAI’s planned campaigns include high-temperature superconductors and an Atlas of the Machine-Accessible Frontier, intended to map where AI can already make meaningful scientific advances.

A company can set the tempo without owning each theorem. It chooses which areas get concentrated resources and which successes get a global launch.

Democratized discovery and one company operating the best private mathematician lead to radically different scientific cultures. I want more labs, universities and public institutions able to run systems at Astra’s level. Otherwise, access to mathematical discovery starts looking suspiciously like an API pricing page.

By 2029, solving another Erdős problem with AI will be routine enough that the announcement struggles to hold the tech news cycle for a full day. Human institutions will be staring at thousands of valid results, each demanding explanation and scarce expert attention.

For centuries, a proof was the finished product. Astra may turn it into the ticket you take before joining the queue.

Frequently asked questions

How much did OpenAI say Astra’s ten mathematical results cost?

OpenAI said the tokens needed to find solutions for Astra’s ten mathematical results would cost roughly $2,000 at Sol API rates. That estimate covers marginal inference for successful solution generation, not model training, researcher salaries, infrastructure, failed experiments, workflow development or debugging.

Do Lean certificates prove Astra solved the original math problems?

Lean certificates establish that encoded conclusions follow from encoded definitions and assumptions under Lean’s logical rules. Reviewers must still verify that those definitions and formal statements accurately represent the historic conjectures, inspect the interpretation and determine whether the resulting methods contribute useful mathematics.

Why could cheap AI-generated proofs create a bottleneck for mathematicians?

Cheap proof generation can produce results faster than specialists can review, interpret and connect them to existing mathematics. Human experts must check formal statements, understand mechanisms, identify reusable ideas, select important results and formulate worthwhile new questions, making qualified attention the scarce resource.