OpenAI's Astra Model Solved 10 Open Math Problems for $2,000
OpenAI's unreleased Astra model reportedly solved 10 decades-old open math problems for under $2,000, including a sphere-packing bound unmoved since 1978.
What did OpenAI’s Astra model actually do?
An unreleased OpenAI model referred to as Astra reportedly produced solutions to 10 longstanding open problems in mathematics, some of which had gone unsolved for decades, at a total inference cost of under $2,000. The results were shared by Noam Brown, an OpenAI researcher who previously worked on the Cicero diplomacy AI and the “o series” of reasoning models. Brown described the results as a potential “major step for scientific reasoning.” Among the problems tackled was a high-dimensional sphere packing bound that had not moved in roughly 48 years, and a question about the existence of “nonsofic groups,” a topic from abstract algebra and group theory.
TL;DR
- Astra, an unreleased OpenAI model believed to sit above GPT-5.1 in capability, reportedly generated solutions to 10 open math problems, several of which had resisted progress for decades.
- The entire run reportedly cost under $2,000, a striking figure given that no human mathematician or research team has ever produced 10 such results in a single lifetime.
- One solved problem involved high-dimensional sphere packing, a question tied to how efficiently data can be encoded and transmitted, and the bound had reportedly been stuck since the late 1970s.
- Astra also reportedly found a counterexample proving nonsofic groups exist, resolving a standing question about whether certain infinite mathematical structures can always be approximated by finite ones.
- Researchers and commentators, including Elliot Glazer, who helped build the Frontier Math benchmark, reacted with surprise, and math-focused online communities reportedly lit up in response.
- OpenAI reportedly said it also tried Astra on Millennium Prize-level problems without success, suggesting current limits, but noted that test-time compute (how much reasoning effort is spent per problem) could likely be pushed further.
- The results reinforce a pattern from earlier OpenAI math work: models appear to advance the field less by inventing alien new math and more by recombining existing ideas across disciplines that few individual humans master simultaneously.
One coffee. One working app.
You bring the idea. Remy manages the project.
Why does solving decades-old math problems matter?
Open math problems that persist for 10, 20, or 50 years are not obscure trivia. They typically mark the edge of what an entire field understands, tested against generations of mathematicians and, in recent decades, considerable computing power. When a problem hasn’t moved in nearly five decades despite supercomputers and sustained expert attention, that stagnation itself is a data point: raw computation alone wasn’t the bottleneck. Something else was missing, whether that’s a new proof strategy, a bridge between subfields, or a person able to hold multiple specialized frameworks in their head at once.
That’s what makes the sphere packing result notable. The high-dimensional sphere packing problem asks how densely you can pack spheres into a space with many more than three dimensions. It sounds abstract, but it connects directly to real infrastructure: encoding data as points in high-dimensional space, spaced far enough apart that noise and interference don’t cause one signal to be mistaken for another, is the same mathematical structure used in error-correcting codes and digital communication systems, including things like 5G. A tighter, better-understood packing bound can translate into more efficient, more reliable data transmission.
According to the reporting on Brown’s post, the constant Astra found for the sphere-packing bound took a notably clean form, expressed in terms of e (Euler’s number) and pi. A clean, elegant constant is often a signal to mathematicians that a result reflects real structure rather than a lucky numerical coincidence.
How did Astra reportedly solve these problems?
The framing from Brown and others is important: this doesn’t appear to be brute-force search. These aren’t problems where throwing enough raw compute at trial-and-error eventually stumbles onto an answer, which is why supercomputers hadn’t cracked them despite decades of trying. Sphere packing and group theory problems at this level aren’t solvable by exhaustively checking possibilities. They require a proof, a logical argument that holds for every case, not just the ones tested.
The more plausible explanation, echoed in the video’s discussion, is that Astra worked by combining knowledge across subfields that rarely get fused by a single human expert. Math and adjacent sciences are deep enough that specialists in one narrow area often aren’t equally expert in another. A model trained across enormous swaths of mathematical literature can potentially notice a connection between, say, a technique from one branch of algebra and a problem in a seemingly unrelated area, then apply it.
This mirrors commentary made about an earlier OpenAI math result, where a respected mathematician reportedly suggested that if you gathered the right group of human experts from different specialties in one room and had them compare notes, they might have arrived at a similar answer. The implication isn’t that the AI invented entirely new mathematics from nothing. It’s that it performed a kind of large-scale synthesis across fields that humans rarely combine, largely because no one person has deep fluency in all of them at once.
What is the nonsofic groups result about?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
The second widely discussed result involves “sofic groups,” a concept from group theory. The general idea: certain infinite mathematical structures can sometimes be closely approximated by finite ones, similar to how you might try to represent an infinitely large, infinitely shuffled deck of cards using only a limited, finite collection of decks. For years, every group mathematicians tested turned out to be “sofic,” meaning it could be approximated this way, which made it tempting to assume this might always be true.
Astra reportedly produced a counterexample: a specific group that cannot be approximated by any finite set, proving that nonsofic groups exist. That result closes a standing open question and, more importantly, demonstrates that the model isn’t just solving numeric optimization problems like sphere packing. It’s also handling abstract, proof-based questions in pure algebra.
Who is Astra, and is it actually GPT-6?
Astra hasn’t been officially named or released as a product. Based on the discussion around Brown’s post, it appears to be a new, more capable model class sitting above OpenAI’s existing GPT-5.x line, described as a possible counterpart to models like Anthropic’s rumored higher-tier releases. Whether Astra is an early codename for what eventually becomes “GPT-6” is speculation at this point, not confirmed by OpenAI.
Elliot Glazer, a set theorist who helped build Frontier Math (a benchmark suite specifically designed to test AI systems against extremely advanced, research-level math problems), was among those who reacted to the results. Frontier Math exists precisely because standard benchmarks became too easy for frontier models. That a researcher closely tied to measuring AI math capability found this result striking is itself a signal of how far the frontier has moved.
Is this actually a big deal, or is it overstated?
It’s worth being precise about limits. According to Brown’s own account, OpenAI also tested Astra against Millennium Prize Problems, the seven (six now unsolved) problems carrying a $1 million reward each for a correct proof, and reportedly had no success there. Those problems remain firmly unsolved. So Astra is not casually clearing math’s hardest known open questions across the board.
What is genuinely new is the combination of scale and cost. No individual human mathematician has ever published 10 comparable results in one career, let alone in one run. And the fact that it reportedly cost under $2,000 in inference spending, not months of grant-funded research or a specialized supercomputer, changes who can access this kind of capability. That price point sits closer to a hobbyist’s API bill than an institutional research budget. Brown reportedly noted that test-time compute (how much reasoning effort the model spends per problem before answering) could likely be scaled up further, suggesting the $2,000 figure represents a floor, not a ceiling, on what’s achievable.
Frequently Asked Questions
What is OpenAI’s Astra model?
Astra is an unreleased OpenAI model, discussed publicly by researcher Noam Brown, believed to be more capable than OpenAI’s current GPT-5.x models. It has not been officially released or fully detailed by OpenAI.
What math problems did Astra solve?
Reported results include a high-dimensional sphere packing bound unmoved for about 48 years and a proof that nonsofic groups exist (a counterexample showing certain infinite group structures cannot be approximated by finite ones), among 10 total problems described as longstanding open questions.
How much did it cost to generate these results?
According to Brown’s account, all 10 results were produced for under $2,000 in inference costs.
Did Astra solve the Millennium Prize Problems?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
No. Brown reportedly said the model was tested against major unsolved problems, including Millennium Prize-level questions, without success.
Is this proof that AI can now do original mathematical research?
It suggests AI models can synthesize existing mathematical knowledge across subfields to produce novel proofs on hard, previously stuck problems. It does not mean AI has solved mathematics broadly or replaced mathematicians, given the failures on the very hardest known open problems.