OpenAI's Newest Model Solved Math Problems Open Since 1978. It Still Can't Tell Your Customer Where Their Order Is.
· CX Pulse
OpenAI's Astra solved ten problems open for decades, for about $2,000 in tokens. That same class of model cannot tell a customer where their order is.
OpenAI's Newest Model Solved Math Problems Open Since 1978. It Still Can't Tell Your Customer Where Their Order Is.
On August 2, OpenAI published ten results produced by an internal version of Astra, its next major model. Each problem had sat open for at least a decade. One of them, the first explicit construction of a non-sofic group, had been open since 1999. Another improved the general upper bound on high-dimensional sphere packing for the first time since 1978.
The token cost of finding all ten was roughly $2,000.
That number is the part worth sitting with, and not for the reason it is being shared. Two thousand dollars of compute produced work that a field of specialists could not produce in forty-seven years. Whatever ceiling people imagined on machine reasoning, it moved this week.
Now go ask it where an order is
That same class of model, dropped into a support queue, cannot tell a customer why their delivery is late.
Not because the question is hard. Because the answer lives in a shipping system it has no access to, under an order number it was never given, tied to an account it cannot verify. The model is not failing at reasoning. It is answering a question about facts it does not have.
This is the gap almost every AI support deployment falls into, and it gets misdiagnosed constantly. A bot gives a vague answer, and the conclusion is that the model is not good enough yet. So the team waits for the next version, or pays for a bigger one, and the vague answers continue.
What actually decides whether it works
Three things, none of which are the model:
What it can see. Order status, account history, the last three tickets, current stock, the real return policy. Not a help article about the return policy. The policy, as your system holds it today.
What it is allowed to do. Issue a refund under a threshold, reschedule a delivery, cancel an order, escalate. An agent that can only talk is a search box with better manners.
What happens when it is wrong. Whether a customer can undo it, reach a person, or flag the answer without starting over.
Every one of those is plumbing and policy. None of them get better when a smarter model ships.
The uncomfortable version
If your support automation is underperforming, the model is probably not the reason. A system that can prove a forty-seven-year-old bound in high-dimensional geometry is not being defeated by a question about a package.
It is being defeated by an integration nobody finished, a permission nobody granted, and a knowledge base last updated in March.
That is a harder problem to fix than swapping a model, which is exactly why the model keeps getting blamed. Buying the next tier is a purchase order. Connecting the order system is a project.
What to check this week
Open your own support bot and ask it something only your systems would know. Where is order 4471. Is the blue one in stock in the Cincinnati store. Why was I charged twice.
If it answers with a policy instead of a fact, you have found your bottleneck, and it is not intelligence.
The frontier moved this week by a distance that is genuinely hard to describe. The gap between that and your queue is not a capability gap. It is a wiring gap, and no one is going to ship a model that closes it for you.
Source: OpenAI, "Ten advances in mathematics and theoretical computer science", published August 2, 2026.