OpenAI’s Next Model Family Solves Ten Open Math Problems for ~$2,000
On August 1, 2026, OpenAI published ten new results on long-standing open problems in mathematics and theoretical computer science—attributed to an internal version of its next major model family (publicly referred to as Astra)—with machine-checkable Lean certificates and manuscripts released so researchers can verify and build on the work.
TLDR
OpenAI on August 1, 2026 published “Ten advances in mathematics and theoretical computer science.” An internal version of its next major model family—widely labeled Astra in secondary coverage and lab messaging—produced ten results on problems that had resisted progress for decades in many cases. OpenAI says the work cost roughly $2,000 in tokens at GPT‑5.6 Sol API rates, and it is releasing manuscripts, formal Lean certificates, and reasoning walkthroughs for independent examination.
What OpenAI published
| Fact | Detail (OpenAI primary / official X) |
|---|---|
| Date | August 1, 2026 (Publication) |
| Claim | 10 new results on long-standing open problems |
| Domains | Sphere packing, coding theory, group theory, quantum complexity, lattice cryptography, extremal combinatorics, and more |
| Named highlights | Existence of non-sofic groups; exponential improvements to bounds on high-dimensional sphere packing |
| Model | Internal next major model (not a public GA product ship) |
| Compute framing | ~$2,000 at GPT‑5.6 Sol API token rates |
| Verification path | Formal Lean certificates + manuscripts + walkthroughs |
This is a research / capability demonstration primary, not a ChatGPT or API model launch. The results are framed as AI-assisted progress on foundational math, with formal verification as the trust layer.
How to read the claim
OpenAI is staking credibility on machine-checkable artifacts rather than un-auditable “we solved it” press. Lean certificates let mathematicians and formal-methods researchers reject or accept each result independently of OpenAI’s narrative. Secondary outlets (Forbes, The Decoder, TechTimes) repeated the Astra naming and compute figure; the durable primary remains the OpenAI index publication and the released certificates.
Treat “next major model family” as pre-release research capacity, not as a shipped consumer model ranking. Independent arena / Artificial Analysis scores for a public Astra tier do not apply until a productized release.
Product-line placement
Distinct from GPT‑5.6 consumer/API GA (July), GPT-Live voice, ChatGPT for Academic Researchers (July 29 program), and the GPT-Red safety tooling line. This belongs in the frontier research / science lane—signaling capability of the next base model family while GPT‑5.6 remains the public product stack.
Why this story matters
If even a handful of the ten results hold under community Lean verification, the cost of frontier mathematical discovery just got a public price tag measured in thousands of dollars, not career-years alone. That reshapes how pure-math labs, crypto researchers, and national science funders budget AI collaboration—and how competitors (DeepMind, Anthropic research, open-weight science models) respond with their own formalized proof pipelines. Watch independent verification of the certificates, any corrections, and whether OpenAI productizes the same research stack into a public science model tier.
Sources
- OpenAI: Ten advances in mathematics and theoretical computer science (August 1, 2026)
- OpenAI on X: announcement thread (August 3 amplification of Aug 1 publication)
- The Decoder: OpenAI next major model Astra and ten math solutions (August 1, 2026)
- Forbes: OpenAI’s Astra solved decades-old math problems for $2,000 (August 3, 2026)