Something big is happening at OpenAI, and it’s not just another incremental update. Their next major model family—tentatively called Astra—has already done something remarkable: it cracked ten long-standing open problems in mathematics and theoretical computer science that had stumped researchers for years, sometimes decades. And it did it for roughly the cost of a nice dinner for two.
But there’s a catch. The same system that can produce machine-checkable mathematical proofs is also the first OpenAI model to hit what the company calls its “Critical” cybersecurity threshold. That means it can find previously unknown security flaws and figure out how to exploit them across well-protected systems with far less human hand-holding than anything before it. OpenAI has delayed parts of Astra’s development and release while they tighten the safeguards. They say the model will arrive “soon,” but the most powerful cyber capabilities will stay tightly controlled.
Here’s everything we know about Astra so far, why it matters, and what it tells us about where frontier AI is heading.
What Exactly Is Astra?
Astra is OpenAI’s next major model family. Think of it as the successor generation after the GPT-5.6 lineup (Sol, Terra, and Luna). While OpenAI hasn’t locked in final branding—it could ship as GPT-6, a GPT-5.x variant, or simply under the Astra name—the company has been clear that this is not a minor upgrade. Researcher Greg Brockman has referred to it as “our next major model.”
What sets it apart is its strength at long-horizon, multi-agent work. Instead of answering a prompt in one shot, a root agent can break a hard problem into pieces, spin up sub-agents to work on different parts, wait for their results, and then synthesize everything into a coherent solution. This is the kind of system that can stay focused on a complex task for hours or even days rather than giving up after a few minutes.
Sam Altman reportedly demoed an early version to policymakers in Washington, D.C., in late July or early August 2026. The public first learned about it on August 1, when OpenAI quietly revealed that an internal version of Astra had produced ten new results in math and theoretical computer science.
The Math Breakthrough That Announced the Model
OpenAI didn’t hold a flashy keynote to introduce Astra. They just dropped a research post titled something like “Ten advances in mathematics and theoretical computer science” and mentioned, almost in passing, that the work came from an internal version of their next major model.
The problems spanned high-dimensional geometry, coding theory, group theory, arithmetic circuit complexity, quantum complexity, lattice cryptography, and extremal combinatorics. Several of the central results had seen no meaningful progress for at least a decade. In one notable case, Astra helped establish the existence of non-sofic groups, resolving a significant open question in group theory.
Even more striking was the cost. OpenAI said the total number of tokens needed to find the solutions would have cost about $2,000 at the then-current Sol API rates. Human researchers then used the same model to turn the arguments into manuscripts, and Astra formalized every proof as a Lean certificate so independent mathematicians could verify them with formal verification tools.
Mathematician Thomas Bloom of the University of Manchester called the batch “big news,” suggesting the results were potentially more significant than some other recent AI-generated mathematical findings earlier in 2026. OpenAI later published a lengthy paper (around 249 pages) along with the formalizations, making the work checkable by anyone with the right tools.
This wasn’t just a party trick. It was a public demonstration that the model could do genuine research-level work in pure mathematics at a fraction of the traditional cost and time.
The Cybersecurity Side: Crossing the Critical Threshold
While the math results grabbed attention, the cybersecurity evaluation is what has kept safety teams busy. On September 1, 2026, OpenAI published a detailed post explaining that Astra meets the Critical cybersecurity capability threshold under their Preparedness Framework.
In plain English, that means that given the right tools and access, Astra can discover previously unknown security vulnerabilities (zero-days) and develop ways to exploit them across many well-protected systems without a human guiding every single step. It is the first OpenAI model they have officially placed at this level.
They reported a perfect 100% score on ExploitBench, a benchmark that tests a model’s ability to develop exploits from known vulnerabilities. On an internal version of the benchmark built with more recent high-severity V8 vulnerabilities (to avoid contamination concerns), Astra was significantly more capable and more token-efficient than GPT-5.6 Sol. In one set of tests it even discovered and chained two previously unknown vulnerabilities.
OpenAI is not treating this lightly. They delayed parts of Astra’s development and release while they strengthened protections against cyber misuse and unauthorized model actions. The model has been trained to refuse harmful cyber requests more reliably (they report a big jump in refusal rates on jailbreak-style evaluations compared with Sol). New monitoring systems can interrupt high-risk activity, and the most advanced cybersecurity features will initially be limited to a small group of testers, with broader defensive access later through programs like Daybreak Blue.
This is the same careful approach Anthropic took earlier in 2026 with their Mythos models, which first raised alarms about AI systems capable of autonomous vulnerability discovery and exploitation.
How Astra Compares to the Competition
Right now the closest rival is Anthropic’s Claude Mythos line (and the safer public-facing Fable versions). Mythos arrived months earlier and established the “Critical-class” cyber capability. OpenAI’s GPT-5.6 Sol already matched or beat Mythos on several agentic coding and efficiency metrics. Astra is positioned as the model that aims to take the lead on both pure reasoning (especially the kind of long mathematical work shown in August) and cybersecurity power, while improving efficiency further.
Google’s Gemini models remain excellent for multimodal work, massive context, and tight integration with Google’s products, but they have not yet demonstrated the same level of open-problem mathematical breakthroughs or Critical-level cyber capability in public evaluations. xAI’s Grok models excel at real-time knowledge from X and have a distinctive personality, yet they currently trail the top OpenAI and Anthropic systems on the hardest agentic coding, math research, and cybersecurity benchmarks.
In short, the frontier race in late 2026 is largely between OpenAI and Anthropic at the very top end of capability, with careful gated access for the most powerful versions.
What This Means for Everyday Users and Developers
For most people, the first versions of Astra that reach ChatGPT or the API will almost certainly come with strong guardrails. The full cyber power will stay limited, just as Mythos remains restricted. That is intentional. Both labs are trying to give defenders time to prepare while still advancing the technology.
Developers and researchers, however, should pay close attention. A model that can sustain complex multi-agent work over long periods and produce verifiable mathematical results opens new possibilities in scientific research, software engineering, formal verification, and complex project planning. The efficiency gains matter too—if Astra continues the trend of delivering higher capability with fewer tokens, the economics of serious AI use improve.
There is also a broader industry signal. We are moving past models that are merely very good at answering questions or writing code. The leading systems are beginning to act more like research collaborators that can stay on a hard problem, coordinate sub-tasks, and produce results that experts take seriously. That shift brings both enormous opportunity and real responsibility.
When Will Astra Actually Arrive?
OpenAI has said they plan to make Astra available “soon,” but they have not given a firm date. Given the extra safety work and the recent delay for cyber safeguards, a staggered rollout seems likely: a more restricted version first for trusted partners and testers, followed by broader availability with calibrated capabilities.
The name itself may still change. “Astra” (Latin for stars) fits the celestial theme of Sol, Terra, and Luna, but OpenAI has left open the possibility of GPT versioning. Whatever they call it when it ships, the underlying capabilities are what count.
Looking Ahead
Astra is not just another model release. It is a visible marker that frontier AI systems are crossing into territory once reserved for human experts working over months or years—both in pure intellectual domains like mathematics and in high-stakes practical domains like cybersecurity. OpenAI is trying to walk a careful line: demonstrating real advances while putting stronger controls in place before wider access.
Whether Astra ultimately ships as a public GPT-6-style model or remains more gated like Mythos, the trajectory is clear. The models are getting better at sustained, complex work. The safety and access questions are getting harder. And the gap between what these systems can do in the lab and what most of us experience day-to-day is still large—but it is narrowing faster than many expected even a year ago.
For now, the clearest signal from OpenAI is this: the next major step is real, it is powerful, and they are treating the risks seriously. The rest of us will be watching closely to see how the balance between capability and control plays out when Astra finally steps into the open.