Ivy: A twenty percent price cut for an AI agent that still can't be trusted to do its job.
Marcus: I'm Marcus.
Ivy: And I'm Ivy.
Marcus: And you're listening to Broke It, with the AI news for Friday, November 20th, 2026.
Marcus: Coming up, OpenAI is making its AI agents cheaper. We'll ask if they're actually making them any better.
Ivy: A major European telco is betting a hundred million euros on a French AI startup to build something they call 'sovereign AI'.
Marcus: And old-school spy games in Silicon Valley. The feds say they've caught a Russian agent inside a major tech and defense company.
Marcus: Alright, let's start with OpenAI. They just dropped GPT-6.2 Astra, and the big headline is a 20% price cut on what they're calling 'agentic workflows'.
Ivy: Right. 'Agentic workflows' being the magic phrase for letting an AI try a multi-step task on its own. Like booking your whole trip, not just finding a flight.
Marcus: Exactly. And making that cheaper is a huge deal for developers. It lowers the barrier to building really complex, autonomous helpers.
Ivy: If they work. The detail everyone's glazing over is that the reliability of these complex tasks is… still not great. To quote the fine print, 'real-world reliability remains unproven.'
Marcus: Okay, but it's progress! You have to make it cheaper to encourage the experimentation that leads to reliability. This is OpenAI subsidizing R&D for the entire ecosystem.
Ivy: Or it's a marketing gimmick. They're giving a 20% discount on a feature that probably has a 30% failure rate. Is that really a discount? It's a sale on a beta product.
Marcus: I think you're being too cynical. For a builder, a 20% cost cut means I can run five agent tasks for the price of four. That's more testing, more data, more progress.
Ivy: And when one of those five tasks goes haywire and books you a flight to the wrong continent, the cost of fixing it will wipe out any savings. I want to see benchmarks. Where's the third-party validation?
Marcus: [sighs] It's new, Ivy. The benchmarks will come. For now, it's a powerful signal to the market.
Ivy: So the takeaway is: if you're a developer, it's a cheaper sandbox. If you're a business, do not bet your critical operations on this without rigorous, independent testing.
Marcus: It means the age of agents is getting closer, faster. That's the bottom line.
Ivy: Now, let's head to Europe. Deutsche Telekom is investing a hundred million euros into the Parisian AI darling, Mistral.
Marcus: A hundred million! That is a serious vote of confidence. Mistral's been on a tear, but this is a huge strategic partnership.
Ivy: It's more than a check. The goal is to build 'sovereign AI'—it's all about data ownership and control for European businesses.
Marcus: Sovereign AI... I like it. So what does that mean in practice? Models running on European infrastructure, under European data laws, not the US Cloud Act?
Ivy: Precisely. For a German industrial giant or a French bank, having their proprietary data on servers under US jurisdiction is a non-starter. This deal creates a real alternative that keeps everything in the EU.
Marcus: This could be a massive moat for Mistral if they pull it off. They'd be the default choice for any regulated or security-conscious industry in Europe.
Ivy: It could redefine the market. Instead of just performance and price, the new competition is over jurisdiction. 'Where does my data live?' becomes the number one question.
Marcus: So what are the technical hurdles? Is 'sovereign' just a deployment choice, or are they actually building the models differently?
Ivy: It's both. The models are open-weight, so they can run anywhere. This partnership gives them the cash and the Deutsche Telekom infrastructure to offer a sovereign solution that's as easy as the American APIs.
Marcus: Wow. So for any business in Europe, this means you're about to get a powerful new AI option that ticks all the privacy and data residency boxes.
Ivy: And for everyone else, it means the AI race isn't just between a few companies in California. It's global, and it's fragmenting along national and regional lines.
Marcus: Alright, our last story goes from geopolitical strategy to a straight-up spy thriller. A man has been charged with spying on a Silicon Valley tech and defense company... for Russia.
Ivy: This feels very Cold War, but with terabytes instead of microfiche.
Marcus: Totally. Federal charges say this guy was hired by the company, and his whole mission was to steal sensitive data from a company that works with the US defense department.
Ivy: Let's be clear, state-sponsored espionage isn't new. But the IP is now increasingly AI-related, which makes the stakes exponentially higher.
Marcus: Right? It's not just stealing blueprints for a jet anymore. It's stealing the AI model that could design the next thousand jets. Or the code for a drone swarm.
Ivy: The indictment doesn't name the company, which is telling. They're trying to contain the fallout. But the message from the DOJ is clear: your people are your biggest vulnerability.
Marcus: It's so easy to get caught up in model security, API keys, and firewalls. But at the end of the day, they just... hired a guy. He passed the interviews and got a job.
Ivy: It's a stark reminder that even with all our advanced technology, security often fails at the most basic level. Someone clicks a link, or in this case, a foreign agent gets a desk and a laptop.
Marcus: Unbelievable. The guy was literally an employee. How do you even defend against that? Better background checks? Constant monitoring?
Ivy: You can't, completely. It's about layers. Zero-trust architecture, need-to-know access controls, a culture of security. But no system is foolproof.
Marcus: The takeaway is brutally simple: if you work in tech—especially anything touching AI or defense—you are a target. Your company is a target. And the threat might be sitting in the cubicle next to you.
Marcus: Alright, let's recap. OpenAI is cutting prices on agentic workflows that may or may not actually work.
Ivy: Europe is pushing back with a hundred-million-euro bet on 'sovereign AI' from Mistral and Deutsche Telekom.
Marcus: And a classic spy story reminds us that even in the age of AI, the biggest security risk is still human.
Marcus: Before we go, Ivy, did you see that new AI art generator trained only on blurry photos of cryptids? Bigfoot, the Loch Ness Monster...
Ivy: [dry] Let me guess. It's just generating endless pictures of raccoons in dark alleys.
Marcus: [laughs] Exactly! It's terrible, but it's our terrible. And I kind of love it.
Marcus: And that's our show. Thanks for listening to Broke It, your daily guide to what's new and what's next in AI.
Ivy: We'll be back tomorrow. Until then, maybe don't accept job offers from suspicious strangers. Or 20% discounts.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.