Episode — 2026-04-13

**Episode title:** The PR War, The Open Revolt, and The Model That Stopped Itself

Anthropic says its latest AI is too dangerous to release. And somehow, that's the best marketing campaign of the year so far. Happy Monday.

Right, let's get into it.

So we start the week with a story that isn't really about AI — it's about reputation. The Guardian ran a piece over the weekend called "Too powerful for the public," and it's a pretty sharp look at how Anthropic has managed to become the "responsible" face of AI while simultaneously building the model that the Pentagon used to strike Iran. That model is Mythos — Claude Mythos — which Anthropic announced last week and then conspicuously decided not to release, citing, and I'm quoting here, "overwhelming responsibility."

Now, the reactions split pretty cleanly. Some people — Scott Bessent at the US Treasury, a Reform MP in the UK — took it entirely at face value and started writing urgent letters to governments. Others, including the AI critic Gary Marcus, pointed out that Dario Amodei, Anthropic's CEO, seems to have graduated from the same school of hype as Sam Altman over at OpenAI — just with better editorial relationships. The New Yorker, the Wall Street Journal, Time magazine — Anthropic's media team has been on a run.

Here's the thing, though — it doesn't really matter whether Mythos is genuinely terrifying or beautifully timed PR. The fact that a US company can build a model used to strike another country, then come out of it looking like the adults in the room, tells you something about where we are. The optics of AI safety have become as important as the practice of it.

Worth keeping an eye on: whether Mythos ever actually ships, and in what form. Polymarket is currently pricing a roughly 4% chance of a public release before April 30th. So don't hold your breath.

Meanwhile — and this is a gear change — something genuinely useful happened in the open-source world. Google DeepMind released Gemma 4 a couple of weeks ago, and it's been picking up steam in the developer community over the past few days.

Now, a 31-billion parameter model beating rivals that are ten times its size sounds like the kind of claim you'd file under "press release noise." But the benchmarks here are holding up — Gemma 4 is outperforming Meta's Llama 4 on maths, coding, and reasoning tasks, and it's running under an Apache licence, which means anyone can use it commercially, no strings attached. It also supports multimodal input — images, video, and audio — across all its model sizes, which is unusual for an open model.

Why does this matter? Because the gap between what you can run yourself and what the big closed labs are selling is shrinking fast. If your business has been told it needs to pay for proprietary AI because there's nothing good enough in the open world — that's increasingly not true. And that changes the calculus on cost, on data privacy, on control.

The bit to watch: whether Google actually builds a community around Gemma 4, or whether it just releases the weights and wanders off. One of those leads somewhere interesting. The other is just a benchmark score.

And finally — quick one, but it connects. OpenAI flagged a security issue over the weekend involving a third-party tool. They've said no user data was accessed. But the timing — right in the middle of the Mythos media cycle — is a useful reminder that the AI security conversation isn't just about what these models might do autonomously. It's also about the infrastructure around them: the tools, the integrations, the supply chain. That's where the boring, expensive vulnerabilities tend to live.

Right. So that's your Monday: Anthropic winning the PR war it wasn't supposed to be fighting, Google quietly releasing one of the best open models anyone's built, and a reminder that the security story is wider than the models themselves.

That's your lot. Fresh week, plenty to watch. I'll be back tomorrow. See you then.