You saw the headline: "OpenAI's Astra solved 10 open math problems for $2,000." And then you probably had two conflicting reactions — either this is the biggest thing in AI since reasoning models, or it is a marketing number dressed up in Lean syntax. The gap between those two reactions is exactly what this article closes.
Here is the situation in one paragraph. On August 1, 2026, OpenAI published ten new results in mathematics and theoretical computer science, credited an internal version of Astra — its next major model, still unreleased — with generating the core arguments, estimated the total token cost at roughly $2,000 at Sol API rates, and shipped every proof as a machine-checkable Lean 4 certificate on GitHub. That is the event. What most coverage does not do is explain what the ten results actually mean, why the $2,000 number matters (and what it leaves out), and how far a machine-checked proof gets you before human peer review takes over. This article does that.
Credibility note: everything below is based on OpenAI's official announcement and paper, the public openai/ten-proofs repository, OpenAI's GPT-5.6 page, and first-party reporting such as Yahoo Tech, cross-checked as of August 5, 2026. Astra itself is unreleased and the results have not been peer-reviewed; treat the timeline and capability claims as OpenAI's own statements until the mathematical community weighs in.
What Is OpenAI Astra?
OpenAI Astra is OpenAI's next major model family, and at the time of writing it has not been publicly released. Every claim in the news is about an internal version of Astra that OpenAI used during development to evaluate the model on open research problems — the same evaluation loop that produced the AI disproof of the Erdős unit-distance conjecture back in May.
Two things are easy to confuse, so let's separate them:
- Astra is not a product yet. There is no ChatGPT model picker entry, no API endpoint, no release date OpenAI has committed to. It is the successor in progress to the current GPT-5 family.
- The proofs are public, the model is not. OpenAI released the results, the Lean certificates, and the reasoning walkthroughs, but explicitly not the model weights or a public preview.
That asymmetry — open evidence, closed model — is the single most important fact for anyone reading the coverage. It means the math can be checked by anyone today, while the capability claims rest on OpenAI's word until the model ships and independent benchmarks appear.
The $2,000 Token Run: Why the Number Matters
OpenAI's exact phrasing matters here, because "solved math problems for $2,000" is not literally what happened:
The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.
Three parts to unpack:
It is a token-cost conversion, not a research budget. OpenAI did not spend $2,000. The company is saying: if you bought the raw tokens those solutions consumed at the retail API price of its current reasoning model, GPT-5.6 Sol, the bill would come to roughly $2,000. That is a way of communicating how cheap the inference was — not a line item OpenAI paid.
The cheapness is the actual headline. Ten problems open for a decade or longer, solved in aggregate for what is roughly the cost of a decent laptop, is the thing to sit with. A few years ago, frontier math progress was measured in researcher-years and expensive supercomputer time. Astra's runs turned that economics upside down: the marginal cost of a new mathematical idea collapsed to the price of a long API call.
It is a lower bound on the real cost, not the full picture. The $2,000 excludes the cost of developing and training Astra, the human researchers who turned raw model outputs into the 249-page manuscript, the formalization work in Lean, and the compute behind the internal version that produced the arguments. The number is a floor on inference cost, not a total. Keep that in mind when someone quotes it as "research for the price of a coffee habit."
The rule of thumb: the $2,000 number is honest about inference cost and silent about everything else. Treat it as evidence of how cheap the runs were, not as the full economics of the achievement.
The 10 Results, Explained in Plain Language
Here is the full list as OpenAI published it, with a plain-English reading of what each result claims. This is where most coverage stops at names; the point of this table is to make each one graspable.
| # | Field | The result, in one plain sentence | What was open before |
|---|---|---|---|
| 1 | High-dimensional sphere packing | New upper bounds on how densely equal spheres can be packed in high dimensions, down to the Cohn–Elkies threshold | For decades, the best upper bounds sat above a theoretical floor set by Cohn–Elkies; Astra's bounds reach that floor |
| 2 | Binary and spherical codes | Exponentially stronger upper bounds on the maximum size of binary error-correcting codes at any minimum distance (plus spherical codes) | Code size bounds had not moved by a full exponential factor |
| 3 | Non-sofic groups | A construction of a non-sofic group — a group that cannot be approximated by finite permutation groups | Whether such groups exist was open since Gromov introduced soficity in 1999 |
| 4 | Connes's rigidity conjecture | A counterexample: some groups are not uniquely pinned down by their von Neumann algebras | Longstanding conjecture in operator algebras, disputed for decades |
| 5 | Arithmetic circuit complexity | New lower bounds for computing the permanent, including a formula lower bound of order n⁴/log n | n³-style barriers had resisted progress for years |
| 6 | Quantum parallel repetition | An exponential parallel repetition theorem for general two-player quantum games | A foundational result in classical complexity with no general quantum analog |
| 7 | Closest vector problem (CVP) | Polynomial-factor hardness of approximation for CVP, with consequences for decoding and lattice problems | Central to lattice-based, post-quantum cryptography; approximation hardness was wide open |
| 8 | Ehrhart's volume conjecture | In every dimension, the sharp maximum volume of a convex body whose centroid is its only interior lattice point | The conjecture was proved only in low dimensions |
| 9 | Multicolor Ramsey numbers | A superexponential lower bound for multicolor triangle Ramsey numbers, resolving Erdős problem 183 | Erdős asked in 1973; the growth rate was unknown |
| 10 | Extremal number conjectures | Counterexamples to the compactness and degeneracy conjectures, resolving Erdős problems 146 and 180 | Two conjectures that shaped extremal graph theory |
Digging into three that deserve it
Non-sofic groups (result 3) — the headline result. A "sofic" group is, roughly, a group that can be approximated arbitrarily well by finite permutation groups — the kind of symmetry you can actually compute with. A huge amount of group theory and ergodic theory quietly assumes every group might be sofic, because the property unlocks powerful approximation arguments. Whether a non-sofic group exists at all was one of the central open questions in the field since Gromov posed it in 1999. Astra's construction produces one explicitly, in Lean. If it holds up, it means an entire class of approximation arguments has a boundary, and researchers will have to say which groups they apply to.
Sphere packing (result 1) — reaching the floor. Packing equal spheres in n-dimensional space is easy to describe and brutal to solve: the densest packing is only known up to a few dimensions. A 2003 theorem by Cohn and Elkies set a theoretical upper bound from linear programming, but for decades no construction got close to it in high dimensions. Astra's new upper bounds reach that Cohn–Elkies threshold. The everyday consequence: density bounds that had stood still for years moved by an exponential factor, and the same linear-programming machinery may carry over to coding theory (result 2, which improved in tandem).
Multicolor Ramsey numbers (result 9) — a 53-year-old Erdős problem. Erdős's problem 183 asked about the growth of multicolor triangle Ramsey numbers R_k(3): the smallest n such that any edge-coloring of a complete graph on n vertices with k colors forces a monochromatic triangle. The answer turned out to be superexponential — a lower bound that grows faster than any simple exponential. It is a clean, decades-old number-theory-style question whose resolution makes for a satisfying story even to people who will never touch graph theory.
Lean Certificates: Why Machine-Checked Proofs Change the Game
The most consequential decision in this release is not a theorem — it is the format of the evidence. Every one of the ten results ships as a formal proof in Lean 4 in the public openai/ten-proofs repository, under an Apache-2.0 license, alongside the 249-page manuscript and the reasoning walkthroughs — a narration of how the model thought through each problem.
Why this is a big deal:
- Lean proof is the standard for "you can verify this yourself." A Lean proof is a set of statements the Lean kernel mechanically checks — there is no trusting the author, no skimming for a gap. OpenAI says the repository builds with Lean 4.32.0 and mathlib, and it publishes the exact commands so anyone can rebuild it.
- It separates "the argument is right" from "the model is smart." If a Lean certificate is flawed, it fails to compile — no subjective reading required. That is precisely why the proofs, not the model, are the strongest part of the announcement.
- It leaves room for human judgment anyway. Lean certificates prove the formal statements are derivable. They do not prove the mathematical interpretation is right, that the chosen statements match the informal problems as the community understands them, or that the results are interesting. That is the work the community still owns.
The one-line takeaway: OpenAI gave the math community exactly what it needs to check the claims independently — and the community is the one who gets to say whether the claims are as big as advertised.
What It Means for You: Release Date, Open Source, and the Catch-Up Race
The questions most readers actually have:
Is Astra released or open source? No to both. It is unreleased and closed, and OpenAI has given no release date. The proofs are open (Apache-2.0); the model is not.
Is this a GPT-5.6 successor? Yes in the model-family sense. OpenAI's GPT-5.6 page documents the current shipped generation; Astra is the next major family still in development. The $2,000 figure is denominated in GPT-5.6 Sol API rates precisely because that is the current frontier token price.
What does OpenAI say it is doing with this? Alongside the announcement, OpenAI points to its ChatGPT for Academic Researchers initiative — free access for 100,000 scientists and mathematicians — and to the May disproof of the Erdős unit-distance conjecture, which it says has already inspired follow-up work such as Bloom–Sawin–Schildkraut–Zhelezov on the sum-product conjecture.
Why are some mathematicians uneasy? OpenAI itself flags the tension, citing the Leiden declaration on AI and mathematics, whose signers worry about AI's role in the field. The core dispute is attribution: OpenAI says it helped prepare and formalize the manuscripts and takes responsibility for correctness, while the mathematical arguments were generated by the system. Whether that division of credit is the right one is an open cultural question, not a technical one.
The Caveats: What the Announcement Does Not Claim
If you want to reason about this event like a researcher, hold three caveats at once:
- Not peer-reviewed. "Published" here means posted publicly, not accepted by a journal. Ten results of this depth will take the community months to verify, interpret, and extend — that is normal and healthy, not a red flag.
- The $2,000 is a retail-token conversion, not a research cost, and it excludes training and human labor (see above).
- Capability ≠ validated. The model is unreleased, unbenchmarked on public tests, and its math ability is demonstrated through these proofs — impressive, but not the same as a reproducible public benchmark.
None of these mean the work is weak. They mean the honest position is promising and checkable, not settled. The proofs themselves are the thing you can act on today — clone the repo and build them.
How to Get Close to This Today
If the Astra story makes you want to push a capable model on genuinely hard reasoning — math, proofs, multi-step logic — there is a meaningful difference between OpenAI's closed path and an open one. Astra is unreleased and closed. GLM 5.2 takes the opposite route: it is Zhipu AI's current flagship, open-weight under an MIT license, with a 1M-token context window and a strong tuning focus on reasoning, coding, and multi-step agentic work — and it is available to try right now, not locked behind an internal evaluation loop. For research-style workloads, it is worth seeing how an open model of that scale handles hard prompts before you commit to a closed-model roadmap.
You can try GLM 5.2 free in your browser on glm5.app — no API key required — and run the kind of reasoning-heavy prompts you care about against it directly. For a deeper look at how it performs on research-style tasks, see our guide to GLM 5.2 for research.
FAQ
What is OpenAI Astra?
OpenAI Astra is OpenAI's next major model family, currently unreleased. An internal version produced ten new results in mathematics and theoretical computer science announced on August 1, 2026, with Lean-verified proofs published on GitHub.
When will OpenAI Astra be released?
OpenAI has not announced a release date. Everything published so far concerns an internal development version; the public release timeline is unknown.
Is OpenAI Astra open source?
No. The model is closed and unreleased. Only the artifacts are open: the proofs, the 249-page manuscript, and the reasoning walkthroughs, published under an Apache-2.0 license in the openai/ten-proofs repository.
How much did OpenAI Astra's math run cost?
OpenAI says the tokens needed to find the solutions would cost roughly $2,000 at GPT-5.6 Sol API rates. That is a retail token-cost conversion, not the full research budget — it excludes model training and human work.
Are the ten proofs verified?
Mechanically, yes: each ships as a Lean 4 formal proof that can be rebuilt and checked. That is different from peer review — the mathematical community still needs to verify that the formal statements match the intended problems and assess the results' significance.
Astra vs GPT-5.6 — what is the difference?
GPT-5.6 is OpenAI's currently shipped reasoning model (its API pricing is the basis of the $2,000 estimate). Astra is the next major model family, still in development and unreleased.
What is a non-sofic group?
A group is sofic if it can be approximated arbitrarily well by finite permutation groups. Whether every group is sofic was open since 1999; Astra produced an explicit construction of a group that is not sofic.
Bottom Line
OpenAI Astra, as of August 2026, is an unreleased next major model whose internal version produced ten new results across mathematics and theoretical computer science — including the first construction of a non-sofic group and sphere-packing bounds down to the Cohn–Elkies threshold — for roughly $2,000 in tokens at current API rates, with every proof shipped as a machine-checkable Lean 4 certificate. The proofs are genuinely open and checkable. The capability claims are genuinely not yet settled.
The most useful way to hold this event is not "AI solved math for $2,000" nor "it's all marketing." It is: the marginal cost of a new mathematical idea has collapsed, and the evidence is now machine-verifiable — but the model behind it is still closed. If you want to start working with open, reasoning-capable models on hard problems today rather than wait for a closed release, GLM 5.2 on glm5.app is a free way to put that to the test right now.
Author: GLM 5 Team. Last reviewed August 2026. This article covers OpenAI's announcement as of August 1, 2026; Astra remains unreleased and the results have not been peer-reviewed. Verify current status on OpenAI's official pages before relying on any timeline or capability claim.
Sources
- Ten advances in mathematics and theoretical computer science — OpenAI — Official announcement (August 1, 2026): the ten results, the Astra attribution, and the ~$2,000 token-cost statement.
- openai/ten-proofs on GitHub — Official repository with Lean 4 certificates (Lean 4.32.0, Apache-2.0), the full ten-result list, and build instructions.
- Ten proofs paper (PDF) — The 249-page manuscript as published by OpenAI.
- Reasoning walkthroughs (PDF) — OpenAI's published narration of the model's thinking for each solution.
- GPT-5.6 — OpenAI — Official page for the current shipped model family whose Sol API rates underpin the $2,000 estimate.
- Model disproof of the Erdős unit-distance conjecture — OpenAI — The May 2026 precursor result from the same evaluation loop.
- ChatGPT for Academic Researchers — OpenAI — OpenAI's free-access program for 100,000 scientists and mathematicians, cited in the announcement.
- Leiden declaration on AI and Mathematics — The open declaration on AI's role in mathematics, which OpenAI cites as the context for the attribution debate.
- Lean — The theorem prover used for the formal proofs; reference for what a Lean certificate is.
- Yahoo Tech: OpenAI Astra model solves 10 open math problems for $2,000 — First-party news report on the announcement and repository contents.