Today in AI — Sep 13

[THEO] Welcome back to The Dish. It's Tuesday, September 13th, 2026, and things are moving fast in AI. Mara, it feels like the whole industry just took a collective deep breath and then let out a very loud exhale.

[DR. MARA] Indeed, Theo. We’ve seen some significant developments from the frontier labs, coupled with a renewed public discussion about the pace of AI advancement.

[THEO] Right, so what's the big headline today? Because my feed is just *exploding* with one particular announcement.

[DR. MARA] The most significant news, without a doubt, is OpenAI's release of GPT-6 Astra. They're calling it a frontier model, and the implication is that it represents a substantial leap in capabilities. It’s being framed as potentially ushering in an "AGI era," though of course, that term itself is subject to much debate.

[THEO] "AGI era"—wow. That's a bold claim. What are people saying about it? Is this just another model, or is there something genuinely new here?

[DR. MARA] Well, OpenAI themselves are pointing to some impressive early demonstrations. They've also rolled out specialized versions, like ChatGPT for Financial Services, which integrates financial data with Astra's reasoning capabilities. And they’ve updated their image generation with ChatGPT Images 2.5, which they say is faster and more faithful. But the biggest splash in the discourse is around a claim that a group of agents, using a next-generation OpenAI model, produced a solution to the Navier-Stokes Millennium Prize Problem.

[THEO] Okay, hold on. The Navier-Stokes problem? For listeners who might not be wrestling with fluid dynamics every day, this is one of the big ones, right? Like, mathematically proving how fluids flow. It's notoriously difficult. If true, that's a *huge* deal.

[DR. MARA] It is. The Navier-Stokes equations describe the motion of viscous fluid substances, and finding a global solution for them in three dimensions under certain conditions is a long-standing challenge in mathematics and physics. A rigorous proof has eluded mathematicians for decades, and it carries a million-dollar prize. If an AI system has genuinely made progress on that, it would signal a profound shift in what these models are capable of in fundamental science. However, it's very early, and the details of the proof and its acceptance by the mathematical community will be critical.

[THEO] Absolutely. And while OpenAI is launching what they're calling the "AGI era," Anthropic is taking a slightly different tack, it seems. They just put out a big report about how their models are being misused.

[DR. MARA] That's right. Anthropic released a detailed threat intelligence report. They've been monitoring and disrupting malicious uses of their Claude models over the past eight months. This includes efforts by suspected state-linked actors for cyber espionage, influence operations, and even attempts related to biological research and weapons development.

[THEO] So, on one hand, we have a huge leap in capability, and on the other, a stark reminder of the risks. And it sounds like the industry leaders are taking notice. I'm seeing a lot of discussion around Anthropic CEO Dario Amodei's call to slow down the pace of frontier AI development.

[DR. MARA] Yes, Amodei has publicly urged AI labs to consider slowing down capability gains, citing risks from AI self-improvement and specific incidents. He's proposed a three-step plan, which includes embedding independent evaluators within labs. What's notable is that several other prominent figures, including Elon Musk and OpenAI CEO Sam Altman, have publicly backed this call for a more cautious approach and stronger safeguards. Altman specifically mentioned that pacing the frontier has been a primary topic of discussion at OpenAI recently.

[THEO] So it's not just Anthropic saying this; it's a broader conversation now. It seems like the industry is grappling with both the incredible potential and the very real dangers, all at the same time. It'll be interesting to see how these calls for caution interact with the rapid release of new frontier models. We'll be keeping an eye on it. That's all for Today in AI. When we come back, we're diving into some fascinating new work on…