TL;DR
The short version
An unreleased OpenAI model, Astra, settled ten decade-old math problems overnight and formalized each as a machine-checkable proof, for a total of roughly $2,000. Even PhD mathematicians say they can't tell whether the proofs are real.
The lesson for leaders isn't "AI is smart now." It's that verifiability now decides what gets automated first, and your job shifts from judging output to building the systems that judge it for you. Analysis from Nathaniel Whittemore's The AI Daily Brief.
Built on the 4 Aug 2026 episode of The AI Daily Brief by Nathaniel Whittemore. Figures verified against OpenAI, Forbes, and Artificial Analysis.
The event that actually happened
Last week OpenAI said its next model, Astra, produced arguments for ten long-open math problems and turned each into a machine-checkable proof, for a total cost of roughly $2,000 — about $200 a problem — at Sol API rates. The math is real. That's not the part worth your attention.
The part worth your attention is what happened next. A data scientist who has spent over 10,000 hours studying math said he couldn't understand the proofs, and neither could his PhD friends. So people did the only thing left: they asked a different AI how hard the problems were. It said each one could plausibly anchor a Fields Medal case.
Astra at Sol API rates — verified via OpenAI-linked coverage (Forbes)
That is the actual event. We built something that outruns the people meant to check it, and the check became "ask another model."
Verifiability is the new sorting rule
Astra's trick wasn't just finding proofs. It formalized each one as a Lean certificate. Lean is a proof-assistant language: a computer confirms the logic holds, step by step, without any human understanding the math behind it. The proof is accepted because it passes the check.
That is why math fell first, and it tells you which of your own work falls next.
Some of the hardest quote unquote work in the world is actually prone to automation first, particularly due to its verifiability — math, cyber, and code.
Aaron Levie, Box
When output can be tested objectively, two things get easier at once: training the model gets cleaner reward signals, and running it lets you confirm it worked, at scale, without a human in the loop. A legal clause, a marketing call, a budget target doesn't have that property — the right answer depends on context and often can't be known for months.
So stop sorting your team's work by how hard it looks and start sorting it by how checkable it is. The impressive-but-verifiable tasks are the exposed ones. The fuzzy, judgment-heavy, slow-to-verify tasks are the defended ones — which is uncomfortable, because the checkable work is often the work your best people are proudest of.
You can't judge the output anymore
Here's the operating problem this creates. For most of software, a senior engineer could read a junior's pull request and know if it was good. That relationship is breaking.
When the model produces a proof no one on the team can verify by reading it, "have a smart person review it" stops working. You need a machine check, like Lean's, or you're trusting output you can't assess. And the trust gap is real: one mathematician flagged that some of Astra's solutions looked wrong, warning that an AI is "as likely to produce a crackpot answer as a human, and they are going to be better at BSing when they do."
The overhang is the work
There's a skeptical reading of Astra, and it's the more useful one. Within 24 hours, people reproduced most of the results using the already-public GPT-5.6, given the right hints. As one commenter summarized, the capability overhang of existing models is only getting bigger.
That means the models you already pay for can likely do far more than you're getting out of them. The bottleneck isn't the next release — it's the system around it: the prompts, the scaffolding, the checks, the redesigned process.
The capability overhang is market opportunity and is where a lot of our time in the near future is going to be spent.
Nathaniel Whittemore, The AI Daily Brief
So the minimum viable move isn't to wait for Astra. Pick one verifiable workflow your team owns, point a current model at it, and build the automated check that lets you trust the result. Math got automated first because it could be verified first. The teams that win the next year are the ones who make more of their own work verifiable, on purpose.
