Three things crossed my desk this week that actually matter, not just noise. A new frontier model, a math proof that took an AI eleven days working mostly alone, and a smart glasses fix that is about a year overdue. Let's get into it.
OpenAI Ships GPT-6 Astra and Calls It Close to AGI
OpenAI rolled out GPT-6 Astra on September 3, first as a limited preview for trusted partners, with a wider release to ChatGPT Plus, Pro, Business, and Enterprise users coming over the following days. The company is calling it the most intelligent and aligned model it has ever built. It posted near perfect scores on reasoning benchmarks and beat both OpenAI's own GPT-5.6 Sol and Anthropic's Claude Fable 5 in head to head tests. It is also the first OpenAI model to hit the Critical tier for cybersecurity capability under the company's own safety framework.
Why it matters: this is the first big model release since OpenAI's rough July, when the Hugging Face incident forced the company to slow down and add more guardrails. A model that scores that high on reasoning and also crosses into Critical territory on offensive cyber capability is a big deal no matter how you look at it. It changes what red teams and defenders both have to plan for.
Robert's take: I do not put much stock in "may represent AGI" talk anymore. Every lab says some version of that at every launch now, and it stopped meaning much a while back. What I do care about is the Critical cybersecurity rating. That is not marketing copy, that is OpenAI's own safety team saying this model can do offensive security work at a level they have not shipped before. Watch what guardrails they actually put around API access, not what they say in the announcement. And at ten bucks in and fifty bucks out per million tokens, this thing is priced for enterprises, not for you tinkering on a weekend project.
Claude Spent Eleven Days Proving Fermat's Last Theorem, Mostly Alone
Anthropic says an internal research model built on Claude formalized Andrew Wiles' 1995 proof of Fermat's Last Theorem into Lean, the language mathematicians use for computer checked proofs. It took eleven days, mostly autonomous, and produced thirteen million lines of Lean code and roughly 29,500 intermediate theorems along the way. That makes it the largest formalization file anybody has ever built.
Why it matters: formalizing a proof means a computer can verify every single logical step, no gaps, no hand waving, no trusting that some reviewer caught every mistake. Wiles' original proof took the math community years to fully verify by hand. Having Claude grind through that same territory in eleven days, with a human mostly out of the loop, says something real about how far these models have come on long, tedious, technically precise work. That is a very different skill than writing a good essay or a clever one liner.
Robert's take: this impresses me more than any chatbot demo I have seen this year, and I do not say that lightly. Anybody can look smart in a five minute conversation. Staying correct and on task for eleven straight days on a problem this hard, with thirteen million lines of code that all have to check out, that is a different animal entirely. This is the kind of AI story that will not trend on social media but probably matters more than the ones that do.
Meta Finally Closes the Loophole on Secretly Recording Glasses
Meta is pushing an update to its AI glasses that stops the camera from recording if you cover the little LED light after you have already started filming. Up until now, some people figured out they could hit record and then cover the light, letting them film folks around them without anybody knowing. Outside services even popped up selling ways to defeat the light. Meta says those workarounds are getting shut down along with this fix.
Why it matters: this comes straight out of real complaints about people getting secretly filmed in public by folks wearing these glasses. A recording light only works as a privacy signal if it is actually reliable, and until now it was not. This is Meta trying to patch a trust problem, not just a technical one.
Robert's take: good fix, about a year late. But let's be honest, a tiny LED on the side of somebody's glasses was never going to be a real consent mechanism in a crowded bar or on a busy sidewalk. Most people do not even know what that light means when they see it. Firmware patches help, but the real fix here is making these devices obviously recognizable as cameras, not quietly closing loopholes after the fact. Meta shipped the patch they had to ship. They still have not shipped the one that would actually fix the trust problem.
That is the roundup. A model that might worry cybersecurity folks, a proof that took an AI longer than most people's vacations, and a pair of glasses that finally does what it always should have done. See y'all tomorrow.