OpenAI Astra: 10 Math Results and a Delayed Launch

"10 advances in mathematics and theoretical computer science" is OpenAI's count of what a model called Astra did to a set of problems that had gone more than ten years without recorded progress. Astra is a model you cannot call, cannot price, and cannot reproduce a single result on. The count went out on August 1–2, 2026 in an OpenAI research post; six days later a second OpenAI post placed Astra at Critical on the cybersecurity track of its Preparedness Framework and pushed general availability back.

Two numbers are carrying the whole week. Ten results, and one Critical designation.

The ten is the number that travelled. It is worth pulling apart before it hardens into the thing everyone knows about Astra.

OpenAI is the only party that has watched Astra work

Every one of the ten results was produced inside OpenAI, on a model OpenAI has not released, described in a document OpenAI wrote. That is a different category of claim from a benchmark score. When a lab posts an SWE-bench or GPQA number, the harness is public, the task set is public, and a third party with API access can rerun it and publish a number that differs from the launch chart.

None of that machinery applies here. There is no Astra endpoint, no evaluation harness anyone can point at these problems, and no second lab with a copy of the model. OpenAI has published no price, no context window, and no general-availability date.

The claim spread fast anyway. The @OpenAI post drew more than 8,000 likes and over 1.5 million views. Sam Altman's @sama follow-up drew more than 20,000 likes and 1.2 million views. Both landed across August 7–8, with screenshots of the math write-ups circulating faster than the write-ups themselves. The Decoder, BleepingComputer and Gizmodo each covered it inside the week.

An advance can be a tighter bound or a single special case

Here is the literal reading, for anyone new to how these announcements are worded.

An open problem in mathematics is a precisely stated question that nobody has answered. "No progress for ten years" means no published paper in that window moved the state of knowledge on that specific statement. OpenAI's word is advances, and that word is narrower than solved. An advance can be a full proof. It can also be a proof of a special case, a tighter bound on a quantity, a counterexample that kills one direction, or a construction that clears the way for somebody else. Ten advances across ten problems is substantial work in fields where a decade of silence is ordinary. The word OpenAI chose covers a wide range of outcomes, and each of the ten sits somewhere inside that range.

Mathematics is a good place to make a claim like this, because a proof is checkable by people who did not produce it. Either the argument holds or a referee finds the hole. That checking runs on the timescale of months, and these ten entered public view at the start of August.

The tally itself is also OpenAI's, twice over. OpenAI chose which problems to work on, decided which outputs cleared the bar for the word "advance", and characterised how long each problem had sat without progress. All three decisions move the number, and all three were made by the same party that announced it. The classification that landed on August 7 keeps the model inside OpenAI while specialists work through the write-ups, so the count and the verification of the count are running on very different clocks.

The ten sit inside mathematics and theoretical computer science

Both fields share a property that almost nothing else in your working life has: the answer is verifiable independently of who produced it, by a procedure that does not involve trusting the producer. A proof of a bound in combinatorics is right or wrong. A refactor of your billing service is neither.

So the ten tell you something specific. They say that a frontier model can now hold a formally specified problem for a long stretch of reasoning, search a space of arguments that human specialists have already picked over, and come out with something the specialists judge to be new. Long-horizon reasoning on a fixed, well-posed target is among the tasks these systems have historically been weakest at, and this is evidence about it.

What the ten say about your inbox, your spreadsheet, or your codebase is inference. The bridge from "produced a novel bound in theoretical computer science" to "will write your migration script correctly" is an assumption, not a measurement. Every lab makes that bridge implicitly in a launch post. Readers who bought GPT-4-era math and coding claims and then met the same models on their own messy, underspecified work know how wide the gap can be.

OpenAI classified Astra Critical on cyber and held the launch

The August 7 post is where the practical consequence lives. Under the Preparedness Framework, OpenAI assigns each model a capability level in each of a handful of risk categories. Critical is the top of that scale. For Astra, OpenAI stated two things together: the model reaches Critical on cyber, and general availability is delayed. Astra is the first model OpenAI has placed at Critical on that track.

Read the two announcements together and the sequencing is doing real work. OpenAI described what Astra can do on August 1–2 and restricted access to it on August 7, and the restriction is the fact that governs anything a reader can plan around this month.

It also gives the ten a second meaning. Automated vulnerability discovery is a search problem over a formal specification, structurally closer to a proof search than to prose. A model that advances decade-old problems in theoretical computer science is a model you would expect to be dangerous at finding bugs in real software. OpenAI's own threshold assignment is consistent with that reading.

Access rules and price decide whether the ten reach you

For a reader deciding what to use this month, the answer is unchanged by all of it. ChatGPT on GPT-5.6 Sol, Codex, Claude and Gemini are the models you can actually put a task in front of today. Astra's ten results do not lower a token bill, shorten a latency figure, or ship a feature.

When Astra does arrive, the Critical classification shapes what arrival looks like. A model held at that threshold on cyber is the kind of release that comes with identity verification, gated tiers, usage review and an application form, because a capability judged dangerous in the wrong hands gets answered by controlling whose hands it reaches. Labs have already built that plumbing for other gated releases. Expect the first availability to reach vetted research and enterprise accounts, and expect a self-serve API key to be a later step. Those expectations are this site's inference from how gated releases have worked elsewhere.

Cost follows the shape of the work. A model that spends a long reasoning horizon on a single hard problem burns tokens in proportion to that horizon, so the per-task bill on problems resembling the ten tracks how long the model thinks rather than how long the prompt is. That is inference from how reasoning models bill.

The count of ten holds as long as each of the ten problems was untouched before Astra and each advance is genuinely new. One prior published solution sitting in Astra's training data would take the count from ten to nine.