AI in 15 — June 29, 2026

Sixteen days ago, two of the best AI models on Earth went dark by government order. This weekend, Washington flipped one of them partway back on — for a hand-picked list of companies that defend power grids and pipelines. The other one, the flagship, is still gone. No restore date. Nobody's promising one.

Welcome to AI in 15 for Monday, June twenty-ninth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

We've tracked this export saga all week, Marcus, but today it actually moved — there's a letter, a reprieve, and a line nobody can draw. That's our lead.

Then — OpenAI builds its own chip, taped out in nine months, with a name from your spice rack.

The "burn all the tokens" era is ending, and CFOs are the ones ending it.

And a game theory professor gets out-game-theoried by his own students.

Lead story, Marcus. We know the backstory — the Friday email, the deemed-export rule, both Anthropic models pulled worldwide. What's genuinely new today?

A letter dated June twenty-sixth, Kate, and it's the first real crack in the freeze. Commerce Secretary Howard Lutnick wrote to Anthropic granting a narrow reprieve. Claude Mythos 5 — the security-focused model — can now be deployed without an export license to a defined set of US organizations that operate and defend critical infrastructure. Power, water, pipelines, the grid. So the model that supposedly broke into classified systems is now cleared, specifically, for the people guarding the systems.

And Fable 5? The flagship everyone actually wanted?

Still dark, Kate. Worldwide. No restoration date. And that's the asymmetry worth sitting with. Mythos was the scarier model on paper — it's the one that triggered Senator Warner's "broke into almost all our classified systems in hours" line. Yet Mythos is the one coming back, because there's a clean national-security use case: hand your best defensive tool to your defenders. Fable is the general-purpose flagship — the one a developer in Berlin or Bangalore would build a product on — and it's the one with no path back.

Remind people of the gap in that "broke into our systems" claim, because that's the part that keeps nagging at me.

It should nag, Kate. That dramatic quote came secondhand, through Senator Warner relaying the NSA and Cyber Command. Twelve days later an anonymous official told the AP something much smaller — the model identified vulnerabilities in an authorized test, which, their words, "did not mean the model was able to exploit them." And Anthropic's own version is smaller still: the jailbreak amounted to asking the model to read a codebase and fix the software flaws in it. Finding a bug so you can patch it is the entire job of a security researcher.

So the same capability is either a weapon or a tool depending on who's describing it.

And depending on what week it is, Kate. Anthropic's argument is pointed — they say holding a model to this standard would, quote, "essentially halt all new model deployments for all frontier model providers," because every capable model can do this. That's not corporate spin; it's the structural problem. If "can find a vulnerability" is the bar for a global shutdown, no frontier model clears it.

So what's the actual takeaway for anyone building on this stuff?

That a single cabinet signature can dark the best model on the market overnight, Kate, and that partial compliance is technically impossible — Anthropic can't sort API users by passport in real time, so the only way to obey was to pull it for everyone. If your product runs on a proprietary US frontier model, you now have a supplier who can be switched off by a letter you never see. That risk did not exist a month ago. And it's exactly why the rest of today's show keeps circling the same question — what can't be switched off?

Quick hits. And the first one, Marcus, is OpenAI trying to own more of its own stack. They built a chip. It's called Jalapeño.

It is, Kate, and the spice-rack name aside, this is a serious move. On June twenty-fourth, OpenAI and Broadcom unveiled Jalapeño — OpenAI's first custom AI accelerator. The key word is inference. This chip isn't for training models; it's for running them, for serving answers to users. And it went from initial design to manufacturing tape-out — that's the point where you commit the design to the factory — in roughly nine months. That may be the fastest high-performance chip development cycle ever.

Nine months is fast for a chip?

Absurdly fast, Kate. These things normally take years. And here's the recursive bit — OpenAI says its own models helped design the chip. AI helping build the hardware that runs AI. It's a massive reticle-sized ASIC — a custom chip purpose-built for one job — and they claim performance-per-watt substantially better than current state of the art. First deployment targeted for the end of 2026.

Why does inference specifically matter so much here?

Because inference is where the recurring cost actually lives, Kate. Training a model is a big one-time bill. But serving nine hundred million-plus weekly users — that cost never stops, every single query. So a chip tuned to exactly how OpenAI's models behave when answering goes straight to their margins, and it loosens Nvidia's grip on their wallet. It's the same instinct every major lab has in 2026 — own more of your hardware, depend on Nvidia a little less.

Next one's the business counterweight to all the capability hype, Marcus. The industry had a word — "tokenmaxxing." And it's dying.

It's reversing fast, Kate. For about a year, employers told staff, especially developers, to use as much AI as humanly possible, cost be damned. Burn all the tokens. That party is over. Uber blew through its entire annual AI budget in four months and slapped on spending tiers starting at fifteen hundred dollars a month per user. The CEO of a startup called Lindy moved one hundred percent of his traffic off Claude to the cheaper open-weight DeepSeek. And Ramp — recently valued at forty-four billion — is building tools specifically so CFOs can see what their token bill is actually buying them.

So the honeymoon's over. The finance department showed up.

The finance department always shows up, Kate. And the timing is delicate for the labs. Anthropic hit roughly a forty-seven billion dollar annualized run rate in May — up from about ten billion in revenue last year. OpenAI was pacing near twenty-five billion. Huge numbers. But the analyst Gil Luria put it bluntly: current growth rates are, quote, "the fastest they will ever be," and some of the biggest enterprise customers may start reining in what he called out-of-control token spend.

And this is why those cheap open models aren't just a technical story.

That's the whole connection, Kate. When a CFO starts scrutinizing the invoice, a free open-weight model that's good enough for routine work becomes a budget weapon — and some of them cost a twentieth of the frontier price. Notion's COO described exactly this swing in an interview this week — from "use AI all day, don't think about cost" to "what is all this spending actually doing for my company?" The capabilities keep improving. The willingness to pay any price for them does not.

Let me do quick callbacks on three we've hit hard this week, Marcus, because they're still live but don't need relitigating.

Briefly, Kate. One — GLM-5.2, the Chinese open-weight model that beats Claude on a cyber benchmark and undercuts GPT-5.5 by a sixth on cost. We dug into it Saturday and Sunday. The only thing I'd add today is the irony in sharp relief: the same fortnight Washington froze Fable over cyber fears, a freely downloadable model matched US labs on exactly those cyber tasks — and it's already on a hundred thousand hard drives. No letter can recall it.

Two — GPT-5.6.

The Sol, Terra, Luna family, Kate — launched to roughly twenty pre-approved organizations at the government's request. We covered it. Still no broad release, and the prediction markets have pushed the likely public launch toward end of July. Same story as the lead, really — frontier cyber models shipping to a tiny vetted list first.

And three — Shazeer to OpenAI.

Noam Shazeer, Kate — co-inventor of the Transformer, the architecture under every modern model, leaving Google to lead architecture research at OpenAI. We noted it earlier in the week. The reason it lingers: his last big architectural idea defined the entire era. OpenAI hiring him to chase next-generation architectures is a bet that the next leap won't come from scaling alone. And it's one more data point in Google bleeding its most consequential people.

Good. No padding.

The show's whole instinct, Kate — when there's nothing new, move on.

Okay, palate cleanser, Marcus, and this one's delicious. A game theory professor at Brown got humbled by his own students.

It's almost too perfect, Kate. Roberto Serrano — a game theorist, mind you — teaches welfare economics and social choice. He gave a take-home exam he'd deliberately made harder than usual. Historic averages in that course ran sixty-five to eighty-five percent. This time? Median score of ninety-eight. Forty of eighty-six students scored a perfect one hundred.

On the hard version.

On the hard version, Kate. He found signs of widespread AI use and collaboration, and he's reverting all his classes to in-person, handwritten exams. And the line going around captures it exactly — a game theorist got made into a sucker by students who, once you incentivize them, simply played the optimal strategy: use the model. They did precisely what his own field predicts rational actors do.

He set the incentives. They optimized. That's the course material.

He should give them extra credit, Kate. And there's a more serious quick one in the same vein — a blogger fed his shoulder MRI into Claude Code and got an analysis that questioned his original radiology report. The comment thread is the real story: actual radiologists showed up. Some said the models are generally terrible at reading medical images without the full 3D dataset. Others noted the clinical guidelines do now advise against the treatment his doctor recommended.

So who was right?

That's the unsettling part, Kate — it's genuinely unresolved. It's a clean picture of the trust gap people are living in. You can't fully trust the AI, but you've also learned you can't fully trust a single human expert either. So you end up triangulating between them. That's not a future scenario. That's a Monday afternoon for a lot of people right now.

One to watch tomorrow, Marcus.

Whether Fable 5 comes back online, Kate, and on what terms. Mythos just got its narrow reprieve for infrastructure defenders. The flagship is still dark with no date. How Washington and Anthropic resolve that sets the precedent for whether any US lab can ship a frontier model without an effective government veto. Watch for a letter, or actual traffic — not a stray config string.

Agree, or counter?

Slight counter, Kate. The more honest signal is the GLM-5.2 download counter. Because if the open frontier is already past the line Washington is trying to hold, then whether Fable comes back is almost a side question. The capability's out there either way. The export letter governs one company. It doesn't govern the hard drives.

That's your AI in 15 for today. See you tomorrow.