OpenAI has revealed Astra, an unreleased model built for problems that take hours or days to solve. The reveal arrived with a striking claim attached. An internal version of the model produced ten advances in mathematics and theoretical computer science, and several of those problems had resisted human effort for decades.
The story reaches well beyond academia. One field on that list sits at the core of the encryption protecting your traffic right now.
What OpenAI Says Astra Can Do
OpenAI describes Astra as its next major model. The company built it for long-running workloads rather than quick replies. Most chatbots answer a question and stop there. Astra aims at work that unfolds across many hours.
OpenAI also says Astra lets AI agents collaborate, with each agent handling a separate piece of one larger problem. That design addresses a real weakness. Long tasks break current models, because context drifts and small errors compound over time. So OpenAI built Astra for exactly that gap.
Ten Problems That Sat Still for Decades
OpenAI reported that Astra worked across an unusually broad set of fields. The list covers high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics.
None of the ten problems had seen movement on their central results for at least ten years. Most had waited far longer than that.
The named results read like a tour through hard mathematics. They include the existence of non-sofic groups, a disproof of Connes’s rigidity conjecture, and fresh bounds for high-dimensional sphere packing. The model also resolved several problems first raised by Paul Erdős, the Hungarian mathematician who left behind a long catalogue of open puzzles.
The price tag
OpenAI put a number on the compute involved. The tokens needed to reach these solutions would cost roughly $2,000 at Sol API rates. That figure sounds tiny for work that stalled specialists for a generation. But it counts only the successful runs, and it ignores the enormous research effort behind the model itself.
Why Lattice Cryptography Belongs in This Story
One entry on that list deserves close attention from anyone who values privacy. OpenAI listed lattice cryptography among the areas where Astra made progress. Lattice math underpins the post-quantum encryption standards now rolling out across VPNs, browsers, and messaging apps.
A lattice works like a grid of points stretched across many dimensions. Finding the shortest route between certain points in that grid gets brutally difficult as the dimensions pile up. Post-quantum schemes lean directly on that difficulty.
So does this threaten your encrypted traffic today? No. The company reported progress on open questions inside the field, not attacks against deployed systems. Still, the direction deserves watching. Crypto agility, meaning the ability to swap algorithms quickly when one weakens, has moved from good practice to basic hygiene.
Machine-Checked Proofs Raise the Bar
One detail separates this announcement from routine AI benchmark hype. Human researchers used the same model to write the arguments up as manuscripts. OpenAI reported that Astra then translated every argument into a Lean certificate.
Lean is proof-checking software. It walks through each logical step and rejects anything that fails to follow. A proof surviving that process carries verification from a machine, not just confidence from its author.
That matters for readers used to unverifiable AI marketing. Most model claims rest on benchmark scores that nobody outside the lab can reproduce. Formal verification gives independent experts something concrete to poke at instead.
GPT-5.7, GPT-6, or Another Name Entirely
OpenAI has not settled on how Astra reaches the public. Reports suggest the model could ship as GPT-5.7, GPT-6, or under a completely different label.
A tiered release also looks possible. Under that approach, consumers would get one version while a more capable variant sits behind special approval. Similar gating already exists elsewhere in the industry, so the idea carries some precedent.
What This Means for Everyday Security
Powerful reasoning cuts both ways. A model that chases a mathematical proof for six hours can also probe software for flaws across the same stretch. Attackers would love that capability, and defenders would gain the identical tool.
OpenAI has not released Astra yet, so the practical impact still sits ahead of us. For everyday users, the advice holds steady. Patch your software, use strong unique passwords, and pick a VPN that supports post-quantum key exchange rather than legacy encryption alone.
Final Thoughts
OpenAI has given Astra a loud introduction, and the mathematics behind it carries genuine weight. Machine-verified proofs are hard to dismiss as marketing. Yet nobody outside the company has run the model against their own problems, so healthy skepticism remains sensible.
The lattice cryptography angle is the piece worth tracking. Encryption guards everything from bank logins to private messages, and it works because certain problems stay hard. As long as OpenAI and its competitors keep aiming their strongest models at exactly those problems, the people building privacy tools have every reason to pay attention.
