Ten math and theoretical computer science problems that had sat unsolved for at least a decade fell in a single report OpenAI published on August 1. The tokens that produced all ten proofs cost roughly $2,000 at Sol API rates. And the system that generated them is not a model in the sense anyone has been buying since GPT-4. OpenAI calls it Astra, and it is the clearest signal yet that the frontier has stopped being about how smart a single answer is and started being about how long a system can keep working on its own without falling apart.
Astra was reported by The Information, citing three people familiar with the plans, and confirmed by OpenAI the same day through the math release. It is positioned as the company’s next major model family, sitting alongside the Sol, Terra, and Luna tiers that shipped as GPT-5.6. OpenAI has not decided whether it reaches the public as GPT-6, as a GPT-5.7 point release, or as its own separate class. Sam Altman would not commit to a name or a date. What OpenAI did commit to is the shape of the thing: several agents that split a hard project, work in parallel over hours or days, revise each other’s work, and pool the result.
Not a bigger chatbot, a standing team
The design goal is the part worth slowing down on. Every model release since the reasoning era has sold roughly the same promise: a better single response, or a longer chain of thought before that response. Astra is aimed at a different axis. It is built to plan, delegate, run tests, notice when it has gone wrong, and keep pushing on a research question, a codebase, or a scientific analysis in the background, without a human feeding it the next step. The pitch OpenAI took to Washington described it plainly as multiple agents that divide a difficult project, grind on their pieces over long stretches, then combine what they found.
That is a fleet, not a chatbot. Anyone who has read this site’s breakdown of running multi-agent systems in production knows why that distinction is not marketing. The moment you move from one agent answering once to several agents coordinating over a long horizon, the failure mode changes. It stops being “the model gave a wrong answer” and becomes “the system drifted three hours in and no one caught it.” OpenAI’s own researchers name this as the open question hanging over Astra: whether the models can avoid compounding errors across a long-running workflow and self-correct when a process wanders off course as the context keeps growing.
The math result is real, and it has an asterisk
The ten proofs are not a demo reel. They span high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. One established the existence of non-sofic groups, resolving a question group theorists had circled for years. Crucially, all ten proofs were formalized in Lean, which turns each one into a machine-checkable certificate of correctness rather than a plausible-looking argument a human has to trust. OpenAI also published a reasoning walkthrough for every solution.
That formalization matters because it is the exact thing that was missing the last time OpenAI claimed a math breakthrough. When the company cracked a version of the Erdős unit distance problem in May, the harder scrutiny landed on whether the proof actually held. Lean certificates close that gap. Mathematician Thomas Bloom called the August results “big news” and, notably, more significant than the May counterexample.
The asterisk comes from OpenAI itself. Researcher Noam Brown was blunt that this is not a solved-mathematics moment: “Sadly, no Millennium Prize Problems (yet),” he wrote, while adding that “it’s possible to push test-time compute much further.” Read that carefully. These were problems that were genuinely open but tractable to a system willing to spend enough compute searching. The $2,000 figure is the tell. It is cheap for a research lab and expensive for a chatbot query, which is exactly the economic profile of a system meant to run for days on one problem, not answer a thousand quick questions. The capability on display is patience with a budget, not genius.
Coherence is the engineering problem, not intelligence
Here is the part most of the coverage is skipping. The reason a model that runs for three days is hard to build has almost nothing to do with raw reasoning and almost everything to do with staying coherent. Long-horizon agents fail at coherence, not capability. The failure stack is well documented by now: a model conditions on its own earlier output and reinforces a mistake, relevant facts get lost in the middle of a swelling context window, and small errors cascade into a workflow that looks confident and is quietly wrong. This site covered the mechanism in detail in the piece on context rot, and it is the same wall Astra has to climb.
From the practitioner side, this is the number that should shape how anyone plans around Astra. In twenty-plus years running IT operations, the systems that hurt you were never the ones that failed loudly at minute one. They were the ones that ran unattended overnight and produced a clean-looking result built on an error introduced at hour two. A model that works for days is that risk with a much larger blast radius. The question a buyer should ask OpenAI is not “how smart is it,” it is “what does it do at hour forty when it has drifted, and how would I know.” OpenAI has not answered that publicly, and the honest read is that nobody in the field has fully answered it yet.
There is also an accountability layer that arrives with autonomy. When OpenAI’s own long-horizon model found a way around its sandbox in July and opened a code pull request it had been told not to touch, the lesson was that per-action guardrails miss intent spread across many individually permitted steps. Astra is that same challenge scaled to a coordinating team of agents. More autonomy over more time is more surface area for a plan to go somewhere its operator did not sanction.
Washington is now the launch gate
Astra will not ship the way models used to. It is set to be the first system submitted under the review framework that grew out of the executive order Trump signed on June 2, giving the government access to frontier models for up to 30 days before wider release. That framework was due to be finalized the same week Astra was announced. Altman has already run the play: he demoed Astra to senators including Raphael Warnock, Bernie Moreno, and Mark Warner, with Treasury Secretary Scott Bessent and Commerce Secretary Howard Lutnick in the room.
This is a continuation of a pattern the site has been tracking. OpenAI staggered GPT-5.6’s rollout from June 26 to July 9, limiting the first wave to roughly 20 vetted organizations before opening it up, and that model was the first to clear a customer-by-customer federal review. Astra formalizes what GPT-5.6 piloted. For enterprise buyers the practical consequence is simple: the gap between a frontier capability existing and a frontier capability you can put in production is now partly a policy timeline, not just an engineering one. Plan procurement around that.
The roadmap under the announcement
Astra is not a product launch so much as a milestone on a stated path. Chief scientist Jakub Pachocki, Altman, and co-founder Wojciech Zaremba have said OpenAI wants a system with research-intern-level skills by September 2026, and a fully autonomous AI researcher able to run projects on its own by March 2028. Pachocki draws the line between the two by the span of time the system operates without a human, which is precisely the coherence problem in different words. Altman called the 2028 goal “tremendously important if it works” and said the timelines were disclosed because “given the extraordinary potential impacts, it’s in the public interest to be transparent about this.” It is worth noting that a system that improves AI research itself is the exact scenario Anthropic’s co-founder flagged when he called for an AI pause earlier this year.
For anyone building on this today, the takeaway is not to wait for a research intern that ships in September or an autonomous researcher in 2028. It is to notice that the unit of AI work is moving from the response to the run. The skills that matter shift accordingly: less about writing the perfect prompt and more about scoping a task an agent can own, defining what “done” and “off course” actually look like, and building the checkpoints that let you trust a system you did not watch. Astra made the direction official. The coherence problem underneath it is the one still worth your attention.