If you've ever handed an AI a project that would take a competent human a few days — a research deep-dive, something with a lot of moving parts — and gotten back something that felt rushed, that's the ceiling OpenAI is now saying it's building past.

On August 1, Sam Altman walked into closed-door meetings with US senators and policymakers on Capitol Hill and showed them a model family called Astra. Not a press release. No product access, no confirmed launch date. The architecture is different from anything currently available: instead of one model answering one question at a time, Astra runs multiple AI agents in parallel — think several specialists dividing a hard problem rather than one person fielding every question — and they keep going in the background for hours or even days before pooling what they found.

To prove this wasn't a pitch, OpenAI published its receipts. Solutions to ten mathematics and theoretical computer science problems that had been open for at least a decade. Problem areas: high-dimensional sphere packing, binary and spherical codes, quantum complexity, lattice cryptography, group theory. Among them: a disproof of Connes's rigidity conjecture — an operator-algebra problem that mathematicians had worked on for roughly thirty years.

Operator algebras are abstract mathematical structures. Connes's conjecture asked whether certain kinds of them are forced to behave in a specific way. Experts had worked on it since at least the 1990s. An AI just settled it. That's either impressive or something bigger, depending on how much you want to sit with it.

Source spread

• The Decoder — OpenAI names Astra via ten math solutions [builder] — Technical breakdown of the multi-agent design and the problem domains. Measured on the gap between demo and product.
• Yellow — OpenAI's next model solved 10 open problems [hype] — Covers the mathematical achievements, heavier on milestone framing than skepticism.
• AI Weekly — OpenAI previews Astra to senators [builder] — Reports on the Capitol Hill demo and the 30-day pre-release review framework context.
• Softonic — OpenAI unveils Astra for harder problems [hype] — Consumer-facing angle on what long-horizon AI would mean for ordinary use.

What's real:

• Math proofs are a qualitatively different kind of evidence than benchmark scores. A disproof of Connes's rigidity conjecture is either correct or it isn't — unlike a 82% coding score, it can be independently verified by any mathematician who reads the proof. If those results hold, they are meaningful in a way that transcends AI-product announcements.
• The multi-agent, long-horizon architecture is the right direction. Current AI is largely constrained to single turns. Hard problems don't work that way. If Astra genuinely sustains coherent reasoning over hours, that's a change in kind, not degree.
• The timing on Capitol Hill is deliberate and worth noting. Altman showed Astra to the people who write oversight rules — just before a 30-day pre-release review framework is expected to land. That's positioning, not coincidence.

What deserves a side-eye:

• No launch date, no product access, no API. "We showed it to senators" is not "you can use it." The gap between OpenAI demos and actual availability has been wide before.
• The name "Astra" is tentative. OpenAI hasn't decided whether this launches as GPT-6, GPT-5.7, or a separate class alongside Sol, Terra, and Luna. That's not a naming detail — it signals how finished the product actually is.
• The ten problems were presumably chosen because Astra could solve them. We don't know how many it failed on, or how it performs on the hard tasks builders actually ship.

What to do about it

• Don't adjust any plans based on Astra's launch. No date, no product, no API. Plan around what you can use today.
• Watch for the math peer review. If working mathematicians confirm the Connes disproof and the other results, that's meaningful signal — more reliable than any benchmark. It will be reported when it happens.
• For everyday users: the "AI that works on your problem for days" isn't available yet. When it is, the most useful version will probably look like: describe a complex project, hand it off, come back to a finished draft. Not magic, but a real shift in what you can offload.
• For builders: the long-horizon multi-agent design may change how you think about agent architectures when it ships. Worth forming a view before it lands — "hours or days" changes the cost and latency assumptions in ways that affect what's feasible to build.
• Watch the regulatory angle. The 30-day pre-release review framework — which Astra may go through before launch — could set a template for how frontier models are released going forward. That matters for everyone building on top of them.

Further reading

• The Decoder — OpenAI announces Astra with ten open math solutions — best technical breakdown of the architecture and problem domains
• Yellow — OpenAI's next model solved 10 problems open for a decade — summary of the mathematical achievements
• AI Weekly — OpenAI previews Astra to senators before 30-day review — Capitol Hill demo and regulatory context
• Softonic — OpenAI unveils Astra model family for harder problems — consumer angle on long-horizon AI