AI in 15 — June 28, 2026

Two weeks ago, one email took the two best AI models on Earth offline. Everywhere. They're still dark. And the bet that GPT-5.6 ships to the public this weekend just collapsed from eighty-three percent down to eighteen.

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

And I'm Marcus, your co-host.

Today the story we've been tracking all week finally shows its second act, Marcus. Washington put its hand on the release valve — and now we get to see where the demand actually goes. Spoiler: it's not staying put.

Then — DeepSeek gives away a free speed boost and, more pointedly, shows its homework.

A Chinese open model nobody can switch off beats GPT on coding for a sixth of the price.

And Google buys a piece of A24 — yes, the prestige movie studio — to put AI inside the editing room.

Lead story, Marcus. We covered the mechanics yesterday — the customer-by-customer access, the phone call from Commerce. What's actually new today?

Two new things, Kate, and they both point the same direction. First, the silence. It has now been more than two weeks since that Friday email — five twenty-one p.m. on June twelfth — pulled Claude Fable 5 and Mythos 5 offline for everyone on the planet. And they are still down. Not throttled, not limited — gone. Anthropic complied but said plainly that recalling a commercial product used by hundreds of millions, if you applied that logic industry-wide, would, their words, "essentially halt all new model deployments."

And the trigger for all this was that dramatic "broke into NSA systems in hours" line.

Which is the part everyone should slow down on, Kate. That quote came from Senator Warner relaying the NSA and Cyber Command chief — Mythos "broke into almost all of our classified systems, not in weeks, but in hours." Terrifying. Except the reporter who first published it, Shashank Joshi at The Economist, later said it should not be read literally. It described an authorized red-team test, where Mythos worked alongside other tools, and it identified vulnerabilities rather than independently breaking in. The NSA hasn't confirmed any of it.

So the same week is either a cyberweapon cracking national secrets, or a security tool doing exactly what you'd hire it to do.

Same facts, two stories, Kate. And that's the whole problem with policy built on a walked-back quote. Now here's the second new development. The bet on GPT-5.6 launching publicly — the model Washington asked OpenAI to slow-roll — has cratered. Polymarket had a launch by today at roughly eighty-three percent. It's now around eighteen. Traders no longer think it ships on OpenAI's timeline. They think it ships on Washington's.

And Sam Altman clearly isn't thrilled about that.

He told staff, a little uneasily, "We've made clear to the U.S. government that this is not our preferred long term model," and that he hoped for a broader release "a couple of weeks later." The administration keeps calling this a voluntary thirty-day review, not a license. But Kate, if the Commerce Secretary phones you personally, and access goes out one approved customer at a time, and the market reprices your launch by sixty points — the word "voluntary" is doing an awful lot of work.

So where does that leave anyone building on these models?

Exposed, Kate. That's the real takeaway and the bridge to the rest of the show. If you built your product on a proprietary U.S. frontier model, you just learned your supplier can be switched off by a letter you never see and never get to contest. So the rational move — if you're a developer anywhere on Earth — is to ask: what can't be switched off? And the answer is sitting in our quick hits.

Quick hits. First one, Marcus, and it's pure engineering — but with an edge. DeepSeek released something called DSpark.

And this is the cheapest good news in AI right now, Kate. DSpark isn't a new model — it's a speed upgrade to their existing DeepSeek-V4. It uses a trick called speculative decoding. Picture a fast little junior model that races ahead and guesses several words at once, and the big senior model checks all those guesses in a single pass, instead of grinding out one word at a time. The result: per-user generation up sixty to eighty-five percent on their Flash model. And the upgraded versions are already on Hugging Face, already serving live traffic.

Faster tokens, no new hardware, free. In the middle of a memory and compute crunch, that's not nothing.

It's exactly the move the moment calls for, Kate. And Nvidia shipped its own version of the same idea this week — dFlash, which we covered, up to fifteen times throughput. So the technique isn't unique to DeepSeek. But here's why Hacker News lit up. DeepSeek published a full paper explaining precisely how they did it, plus an MIT-licensed codebase. The top comment contrasted that with American labs that, quote, "no longer do, unfortunately."

Is that fair, though? That feels like a vibe more than a measurement.

It's a vibe, and I'll flag it as one, Kate. These are self-reported benchmarks — DeepSeek grading DeepSeek. "Open and innovative versus closed and benchmark-chasing" is a community narrative, not a measured fact. But the openness itself is real and checkable: a detailed method, MIT code, anyone can run it. And the timing is brutal for the West — a Chinese lab publishing everything in the open, the exact same fortnight U.S. labs are being pulled offline by export law.

Which lands us right on the next one, Marcus. GLM-5.2. This is the model that can't be switched off.

This is the second act made concrete, Kate. Z.ai — formerly Zhipu — released GLM-5.2, a seven-hundred-fifty-three-billion-parameter open-weights model, MIT licensed, built for long-horizon autonomous coding. It scored sixty-two-point-one on SWE-bench Pro, beating GPT-5.5's fifty-eight-point-six. And the price is the headline: roughly a dollar-forty in, four-forty out per million tokens, against GPT-5.5 at five and thirty, or Claude Opus at five and twenty-five.

So a sixth of the cost, and it's already winning on the benchmark that's supposed to be the closed labs' moat.

Agentic coding was the moat, Kate. That was the one thing you couldn't get cheaply anywhere else. And now you can — openly, with a one-million-token context window, already a top-ten most-used model worldwide. The licensing docs are almost taunting: they advertise "no regional limits" and "technical access without borders." That is a direct shot at the export regime. You pulled the best Western models off the global market — fine, here's one nobody can recall, because the weights are already on a hundred thousand hard drives.

And the same standing caveat applies to that benchmark number.

Self-reported, hold it at arm's length, Kate. But there's a sharper point Axios made that I want on the record. The exact openness that makes GLM-5.2 attractive to honest developers also hands capable cyber tooling to anyone — including bad actors. Which is the precise capability concern that got Mythos pulled in the first place. So Washington's lever may have moved market share — toward Chinese open models — faster than it moved safety. The capability didn't disappear. It just changed passports and lost its guardrails.

We should note this is the predictable shape of the whole week's arc.

It genuinely is one story, Kate — government gatekeeping at the top, demand routing down and out to whoever's still shipping. The ban, the GPT-5.6 hold, DeepSeek's openness, GLM-5.2 — that's a single thread. And TechCrunch is tracking a broader wave: Asian startups, including Sakana's "Fugu" system, all rushing out Mythos-like offerings to fill the exact gap the export ban created. Restrict the supply, and the market builds you a new one.

Let's change the channel completely, Marcus — because this next one made me do a double-take. Google is investing in A24. The movie studio.

It surprised me too, Kate. Google is putting roughly seventy-five million dollars into A24 — its first-ever financial stake in a Hollywood studio. A24 is the prestige outfit behind films like "Backrooms" and "Marty Supreme," famous for fiercely protecting artistic vision. The money is tied to a research partnership with DeepMind to build AI tools for filmmakers — things like AI-generated storyboards. DeepMind researchers will actually embed inside A24's production workflows.

And the interesting part is what's not in the deal.

That's the whole tell, Kate. Google does not get A24's content library, and does not get its data for training. It's capability access, not a data sale. Scott Belsky — who runs A24's new tech division, "A24 Labs" — went out of his way to frame this as different from other AI-film deals. He took a swipe at developers who pitched AI mainly as a way to make movies "cheaper and faster."

Which is exactly what artists are terrified of.

Right, and that's the bet, Kate. A24 is trying to prove AI in film can be sold as a creative tool — craft, not cost-cutting — and they may be the one studio with enough credibility with skeptical artists to even attempt that case. Whether "we kept our data, it's about the work" survives contact with a quarterly budget meeting is the open question. But the structure matters more than the press release: access to capability without handing over your rights. Every other studio and rights-holder is going to study that template.

Quick callback, Marcus — two stories we covered yesterday that moved a touch but don't need relitigating.

Two threads, briefly, Kate. Apple's price hikes — Macs and iPads up fifteen to twenty-five percent, blamed squarely on the AI-driven memory shortage. We did that in full Saturday; nothing's changed except Apple stock is still nursing that six-percent drop and DRAM is still forecast up another roughly sixty percent this quarter. The build-out is still on your shopping cart.

And the Codex numbers.

OpenAI's claim that Codex now generates ninety-nine-point-eight percent of its internal output tokens — also Saturday, Kate. Same caveat stands: self-reported, "output tokens" flatters a verbose agent, hold the precision lightly. But the direction — agents going from demo to default, spreading into legal and finance — is the real signal, and it's consistent with everything else we're seeing. I'd just move on.

Good. No need to repeat ourselves.

The show's whole instinct, Kate — when there's nothing new, don't pad it.

One to watch tomorrow, Marcus.

GPT-5.6's actual public release, Kate. Whether OpenAI ships it broadly — and on whose timeline — is the live test of whether this "voluntary" review is a one-off or the new normal. The prediction markets have already pushed the likely launch toward end of July. Watch for any signal of wider rollout, or a new approved partner quietly getting access.

Agree, or counter?

Slight counter, Kate. The more interesting tell isn't OpenAI — it's Anthropic. Whether Fable and Mythos actually get switched back on. Two-plus weeks of total silence on the two best models on Earth is the real signal here, and nobody in Washington is explaining it.

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