1. What OpenAI actually released

On August 1, 2026, OpenAI announced ten results spanning high-dimensional geometry, coding theory, group theory, circuit complexity, quantum complexity, lattice cryptography, convex geometry and Ramsey theory. It attributes the mathematical arguments to an internal version of Astra—described as its next major model—while humans and the same model prepared manuscripts and Lean certificates.

The current 253-page manuscript is marked “Updated August 6, 2026” and links to its original version. This is therefore a revisable research release, not public access to Astra or a final journal record. As of August 17, the OpenAI page does not identify a peer-reviewed venue that has accepted the collection.

2. Ten results do not all mean ten complete solutions

OpenAI's own wording says the collection resolves or makes substantial progress on long-standing problems. It includes a construction of a non-sofic group and a counterexample to Connes's rigidity conjecture, but also improved bounds for sphere packing and codes. The paper calls the sphere-packing exponent its first general improvement since 1978; improving a bound is not the same as classifying every case.

The defensible headline is ten formalized contributions of different types—not “Astra solved ten open problems.” Specialists must still judge the scope, priority and importance of each theorem in its own literature.

  • Resolution or construction: proves an object exists or closes the stated question.
  • Disproof: supplies a counterexample to a conjecture.
  • Bound improvement: advances the best known value without ending every version of a problem.
  • Formal certificate: checks logic in an encoded system; it is not a significance score.

3. What Lean actually guarantees

The repository's All.lean file imports ten modules, one per result, and the project specifies Lean 4.32.0, mathlib and Lake. The Lean Reference Manual says kernel acceptance means a proof of the formal statement follows from the definitions, theorems and axioms in the file and its imports. Proper validation catches unfinished goals and tactic failures.

OpenAI also publishes Comparator challenge instructions for sandboxed rechecking with an external checker. That strengthens assurance by protecting the trusted statement and reducing reliance on one build path. Yet even gold-standard checking cannot answer a semantic question: does the formal statement exactly capture the community's intended conjecture?

4. Four gaps a certificate does not close

First, translation: correct code can encode a weaker claim than the informal headline. Second, novelty: a kernel does not search the literature or detect rediscovery. Third, significance: logical validity does not establish depth or usefulness. Fourth, attribution: a certificate cannot prove that references are complete or that training material was properly licensed.

The Leiden Declaration recommends disclosure, human explanations and formal verification where appropriate, while retaining human responsibility and peer-reviewed publication. That is not an argument against Lean. It places Lean inside a broader chain of scientific trust.

  • Lean asks: is the encoded derivation accepted?
  • The domain reviewer asks: are translation, prior art and novelty sound?
  • The journal asks: did method, attribution and importance survive independent review?
  • The institution asks: can the claim be reproduced without privileged model access?

5. What the $2,000 figure means

OpenAI says the total tokens needed to “find solutions” would cost roughly $2,000 at Sol API rates. That is a token-price conversion, not a project accounting statement. The announcement does not clearly enumerate failed attempts, problem selection, Astra's training compute, expert time, preparation of 253 pages or formalization and review costs.

It is therefore wrong to say ten mathematics problems cost only $2,000 to solve. The genuine economic signal is that marginal inference for successful discoveries may be very cheap relative to researcher time. Access terms, hit rate and human verification will determine the economics of automated mathematics labs.

6. A procurement protocol for research institutions

A university or national lab evaluating an automated-mathematics system should demand held-out problems, buildable certificates, two domain reviewers and a ledger of failed runs, tokens and human hours. Discovery cost must be separated from the cost of turning an idea into a publishable and understandable proof.

A lower-risk starting point is an assistant that searches hypotheses, proposes lemmas and formalizes known steps, with a named human gate before any priority claim. Multilingual institutions should also audit terminology: translating an informal theorem into Arabic, Korean or Japanese is part of knowledge integrity, not packaging after the proof.

  • Rebuild certificates from pinned toolchains and dependencies.
  • Map every formal statement to its prose claim line by line.
  • Run an independent priority search before publicity.
  • Publish negative runs and review costs, not only successful-token cost.
  • Name a human accountable for correctness, references and impact.