OpenAI has named its next major model Astra, and it introduced the model by publishing ten new results in mathematics and theoretical computer science. Each result answers a problem that had stayed open for at least ten years. The company published the work on August 1 and put the total compute bill at roughly $2,000 (around ₱122,500) at the API rates of its current Sol model. Astra itself is not something you can use yet, and OpenAI has not announced a release date.
What OpenAI announced
Astra is described as a model family built for long-running work — tasks that run for hours or days rather than seconds — using several agents that coordinate with each other. An agent here simply means a copy of the model that can plan, run steps, and check its own output. The Decoder reported that OpenAI has not settled on a final product name; the model could ship as GPT-5.7, GPT-6, or something else entirely.
OpenAI framed the ten results as coming from an internal version of Astra, not the public models. The company said the problems it worked on had "seen no progress on their central results for at least a decade, and in most cases, much longer."
The ten results, in plain terms
The work spans high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. The headline items, as listed by The Next Web, include:
- The first explicit construction of a non-sofic group, answering a question the mathematician Mikhail Gromov raised in 1999. Nobody had shown such a group exists — or shown it cannot — in the years since.
- A disproof of Connes's rigidity conjecture, a long-standing claim about a family of mathematical objects called von Neumann algebras.
- A proof of Ehrhart's volume conjecture.
- Solutions to three problems posed by Paul Erdős, including his problem #183 on multicoloured Ramsey numbers.
- A better bound on sphere-packing density in high dimensions — the first improvement on this measure since 1978.
- A parallel repetition theorem for two-player quantum games, and new circuit complexity lower bounds for computing the permanent.
Lattice cryptography appearing on that list is worth a note for security readers: lattice problems are the mathematical base of the encryption standards meant to survive quantum computers. Nothing in the release breaks those standards, but it shows where the model was pointed.
Why the Lean certificates matter
The more important detail is not the list — it is how the work can be checked. OpenAI released a 249-page manuscript, 62 pages of notes narrating how the results were found, and a repository of Lean 4 certificates. Lean is a proof assistant: software that checks a mathematical argument step by step and refuses anything that does not follow. A machine-checkable certificate means a reader does not have to trust the model, or OpenAI. They can run the check themselves.
That matters because large language models are known to produce confident, wrong arguments. A Lean certificate removes that question for the steps it covers.
What it cost, and what that implies
The $2,000 (around ₱122,500) figure is for the tokens used to find all ten solutions, priced at Sol API rates. That is the cost of the successful runs, not the full research effort behind Astra, and it is not a price list for anything you can buy. Still, it is a striking number next to the decades of human effort these problems absorbed. Our earlier coverage of OpenAI's GPT-5.6 Sol rollout explains the model tier that pricing refers to.
How mathematicians are responding
Thomas Bloom of the University of Manchester called the results "big news" and ranked them clearly above earlier AI results in mathematics. He also pushed back on the idea that this replaces mathematicians, pointing out that the system leans on more than a century of built-up theory to get anywhere.
The reception is not uniformly warm. The Next Web notes the tension with parts of the mathematics community over announcing results on a company blog instead of through peer review — a concern raised in the 2024 Leiden Declaration, which warned that AI companies were bypassing normal review and attribution. Fields Medallist Tim Gowers has previously argued that AI-generated proofs of this quality belong in top journals. Whether blog-announced results get that treatment is still unsettled.
BleepingComputer reports that Astra is expected to be the first model submitted under a new US government safety review framework, which would require federal sign-off before a public release. That is one more reason not to expect it soon.
What a Filipino developer or student should take from this
Nothing here is available in the Philippines, or anywhere else, today. Astra has no release date, no pricing, and an extra regulatory step ahead of it. The practical read for local developers is narrower: the models you can already reach through the API — and through consumer bundles like Globe's AI Fiesta — are not the ones that did this work.
The part worth copying is the verification habit. If a model's output has to pass a checker before anyone believes it, the checker is doing the real work. For students and engineers here using AI for proofs, code, or analysis, the lesson is the same: build the check, then trust the output.
