Why everyone thinks OpenAI's Astra model is secretly GPT-6

A model nobody outside OpenAI has ever touched just settled a math question that's been open since 1999. No product page, no waitlist, no API key — just a 249-page manuscript and a name: Astra. Within hours, half the AI world had decided this was GPT-6 wearing a disguise. The other half pointed out that OpenAI never actually said that.

Both sides are working off the same set of facts, which is exactly the problem. This is what OpenAI actually published, why the rumor mill ran so far ahead of it, and how to check any of these claims yourself instead of trusting a screenshot on X.

What did OpenAI actually publish about Astra?

OpenAI announced that Astra, its next major model still awaiting public release, had generated solutions to 10 longstanding problems across mathematics and theoretical computer science, each unsolved for ten or more years. This wasn't a benchmark score or a leaderboard flex — it was a research paper with receipts. Alongside the announcement, OpenAI released a 249-page manuscript and Lean 4 proof certificates on GitHub under an Apache 2.0 license, and the repository's "sorry" count stands at zero, meaning every step across all ten formalized proofs is fully verified.

The problems themselves span some genuinely hard territory. Astra disproved Connes's rigidity conjecture on von Neumann algebras, proved Ehrhart's volume conjecture, resolved three problems from Paul Erdos's catalogue including problem 183 on multicoloured Ramsey numbers, produced the first improvement to the general upper bound on high-dimensional sphere-packing density since 1978, and proved a parallel repetition theorem for two-player quantum games. The standout, according to most mathematicians who weighed in, was the non-sofic groups construction — no mathematician had managed to prove or disprove whether non-sofic groups exist in the 27 years since the concept was introduced.

And the cost is the part that should actually worry other labs more than the proofs themselves. OpenAI put the total compute cost for all 10 solutions at roughly $2,000 at GPT-5.6 Sol API rates. That's not a typo. Ten problems that resisted human mathematicians for decades, cracked for less than a decent office chair.

Is Astra secretly GPT-6?

Here's where the story gets murky, because OpenAI never made that claim. OpenAI has not said when Astra will be released publicly, describing it only as its "next major model," though some observers have speculated that Astra is the GPT-6 series. Separately, other reporting frames it differently: Astra is an unreleased OpenAI model believed to sit above GPT-5.1 in capability.

Notice the gap between those two framings — "next major model" and "sits above GPT-5.1" aren't the same statement as "this is GPT-6." That distinction matters because the internet doesn't do nuance well. A vague internal codename plus a jaw-dropping result is exactly the kind of vacuum that gets filled with confident guesses, leaked-benchmark screenshots, and Discord threads insisting they have insider info. None of those things are OpenAI's official position.

What is confirmed is the reaction from people whose job is to actually check this stuff. The company's head of mathematics research, Sebastien Bubeck, confirmed the results on X, calling them "beautiful." Research scientist Noam Brown called the results "a major step for scientific reasoning" in a post on X, and OpenAI credited Astra with the underlying mathematical reasoning while human researchers worked the model's output into papers suitable for publication. Outside reactions were just as strong — Thomas Bloom, who curates the erdosproblems.com database, described the ten results as "big news" on X, ranking their significance above a previous milestone.

Also worth knowing: this isn't Astra's first appearance. In May the same long-horizon model family disproved the Erdos unit distance conjecture, an 80-year-old problem in discrete geometry, and Fields Medalist Tim Gowers said he would recommend that proof for publication in Annals of Mathematics without hesitation. So this is a pattern, not a one-off stunt.

Why do Lean-verified proofs matter more than a benchmark score?

This is the detail that separates Astra's announcement from the usual "our model beat GPT-5 on MMLU" press release. Lean is a proof assistant, not a vibe check. Lean is a programming language and proof assistant that forces every step of a mathematical argument to be spelled out in machine-readable detail. That means a compiler — not a human grader, not a leaderboard — decides whether the logic actually holds.

But it's worth being precise about what verification does and doesn't cover. Lean's compiler guarantees that the logical steps in a proof are internally consistent and follow correctly from one another, but it doesn't independently confirm that the problem was formalized to mean what mathematicians intended — that's still something human reviewers need to check. In other words, the machine confirms the math is airtight; humans still have to confirm the question being answered is the right question. And the peer-review clock hasn't finished either — the results are Lean-verified and have been reviewed informally by mathematicians who saw preprints, but none have gone through a formal, refereed journal process yet.

If you want to see what Lean verification actually looks like, the official Lean documentation walks through how the proof assistant checks each logical step, and you can browse the community-maintained math library, mathlib, to see the scale of formalized mathematics Lean already supports.

What role does Leanstral play in this story?

One name that got unfairly swept into the "secret GPT-6 leak" rumor pile is Leanstral — and it's actually a real, documented product, not a rumor. Mistral's Leanstral 1.5 is designed to work with Lean 4, an interactive theorem prover used to express and verify mathematical statements and software specifications, targeting both academic mathematics and practical software verification. It uses a Mixture-of-Experts architecture with 119 billion total parameters, though only around 6 billion parameters are active during inference.

The results are genuinely competitive with much larger systems. Despite its size, Leanstral delivers results that rival far larger and proprietary systems — it saturates miniF2F, solves 587 out of 672 problems on PutnamBench, and reaches a new state-of-the-art on FATE-X and FLTEval. The cost story is dramatic too: on PutnamBench, Leanstral edges out a competing high-setting prover by 7 problems at far lower cost — about $4 per problem, against an estimated $300 or more for the alternative.

Mistral publishes full details on the official Leanstral 1.5 announcement page, and the model itself is downloadable through Hugging Face for anyone who wants to run formal verification locally. The lesson here: not every "leaked" name in this story is fake. Some of them are just real products that got dragged into a rumor that had nothing to do with them.

What's happening with Gemini while everyone watches OpenAI?

Google hasn't been sitting still during any of this. Gemini 3.6 Flash is Google's new default Gemini model, released with a 1 million-token context window, a March 2026 knowledge cutoff, and lower per-token pricing than Gemini 3.5 Flash. Compared to 3.5 Flash, it consumes 17% fewer output tokens while taking fewer reasoning steps and tool calls to accomplish multi-step workflows, and it's priced lower at $1.50 per million input tokens and $7.50 per million output tokens.

Notably absent from that release: the flagship Pro model people were expecting. Google did not release Gemini 3.5 Pro in this announcement — the company said the Pro model fell short of internal expectations on coding and complex reasoning, so its broader release was delayed. That's a useful reminder that every lab is quietly shipping incremental wins and holding back the ones that don't hit the bar, which is exactly the kind of context that gets lost when a rumor thread only talks about one company.

You can check Google's official specs and benchmarks directly on the Gemini 3.6 Flash developer documentation rather than relying on secondhand summaries.

How can you compare Astra's claims against Claude and Gemini yourself?

Since Astra isn't publicly available, you can't test it directly yet. But you can absolutely stress-test the surrounding claims — run the same math or reasoning prompt across GPT, Claude, and Gemini and see where they actually diverge, instead of trusting a "leaked benchmark table" that nobody can verify. That's the oldest trick in this genre: a screenshot with suspiciously precise numbers that nobody can check.

How do you set up a side-by-side comparison?

A multi-model chat tool is the fastest way to do this without paying for four separate subscriptions. Helio lets you run prompts across models in one interface, so you can throw the same question at GPT and Claude side by side and compare reasoning styles directly. If you want a broader model roster for pure comparison work, ChatHub supports GPT-5, Claude, Gemini 3, and over 20 other models in one workspace, with a built-in mode for sending one prompt to multiple models at once.

Here's a simple workflow to run before you believe any "GPT-6 destroyed Claude" headline:

1. Pull the actual claim (a specific benchmark number, a specific proof, a specific quote).
2. Search for the primary source — the paper, the GitHub repo, the official blog post — not the tweet summarizing it.
3. Run the same question through two or three models side by side and compare reasoning, not just the final answer.
4. Check whether the result is independently verifiable (like a Lean proof) or just a self-reported score.

Should you treat Astra as a general leap or a specialist tool?

This is the part that gets flattened in every "GPT-6 is here" headline. A model that can grind through a sphere-packing bound for hours with a compiler checking its work is not the same thing as a model that's broadly smarter at everyday tasks. Proof verification gives the model an unambiguous signal of success or failure — something most real-world tasks simply don't offer. Until OpenAI demonstrates this kind of long-horizon reasoning working on problems without a built-in checker, the safer read is that Astra is an extraordinary specialist rather than proof of a general capability leap.

That's not a knock on the achievement — it's genuinely remarkable. It's just a different claim than "OpenAI's next chatbot will destroy Claude at everything," which is the version that spread on social media.

What should you actually do with this information?

Bookmark the primary sources, not the reaction threads. Astra's Lean proofs are public and checkable by anyone with the patience to read them — that's rare enough in AI announcements that it deserves attention on its own merits. Whatever OpenAI eventually calls its next public model, the real story here isn't the name. It's that a machine-checkable proof now costs less than a month of streaming subscriptions, and that changes who gets to attempt the next impossible problem.