Before we get into this week’s AI news, I want to share a quick update on The Blueprint.
Last week I didn’t send a newsletter. That was the first time in nearly three years, and after more than 150 issues, that felt strange. But the honest reason is simple: I didn’t think there was enough genuinely interesting news to share.
This reflects something I’ve been feeling for a while. AI as a topic is changing. For the last three years there has been a constant stream of major model launches, new products, big announcements and chaotic industry drama. The weekly newsletter format made sense because there was a lot to keep up with, and most of it needed explaining quickly.
A month ago I sent out a reader poll asking what you value most, what you’d like more of, and whether the weekly format is still useful. Only a few people responded, so I’m not going to pretend the results are representative of the whole reader base. But they were still helpful, and they broadly confirmed my instinct that the most useful version of The Blueprint is probably less about volume and more about being useful and interesting.
AI is starting to mature. There are fewer genuinely big weekly news moments, and the more interesting questions are becoming more practical and specific. How does AI change marketing? How does it change work? How do businesses actually implement it? What should we do differently now that the technology is moving from a novel technology to broader infrastructure?
So I’m going to change how I approach The Blueprint. I’ll still write about the major AI stories when they matter, but I’m going to move away from trying to publish a weekly round-up every week. Instead, I want to focus more on bigger opinion pieces, sharper points of view and practical analysis when something important happens.
That feels like a better fit for where AI is now, and a better use of The Blueprint. Less “here’s everything that happened this week” and more “here’s what this means, and why I think it matters”.
So this won’t be the end of The Blueprint. But it is the end of The Blueprint as a weekly newsletter.
That makes this week’s issue the final weekly edition. It feels appropriate to end this version of The Blueprint with the kind of stories that made me want to write it in the first place: frontier AI, regulation and what all of this means.
So, for one last time in the weekly format, here’s what mattered this week.
OpenAI will delay GPT-5.6 after Trump administration request
OpenAI is reportedly delaying the full release of GPT-5.6 after a request from the Trump administration, with access initially limited to a small group of enterprise customers. I’m not sure this is what regulation should look like, but the US government is expected to approve customer access on a case-by-case basis during the preview period.
That is a strange place for AI to have arrived. For years, the story has been about moving faster: faster models, faster deployment, faster adoption. Now the question is becoming much more political: who gets access to the most capable systems, and who gets to decide?
You can understand why governments are nervous. These models are getting more capable, and some of the most sensitive use cases sit uncomfortably close to cyber security, national security and economic power. But case-by-case approval also creates a very messy precedent. It turns model access into something that can be negotiated, withheld or granted by the US government.
This is not just a US issue either. If the best models are built by American companies, and the US government starts deciding who can use them, every other country has a problem. Businesses, governments and researchers outside the US may find that access to frontier AI becomes a political question too.
Inside the fight over Claude Mythos 5
Two weeks ago I wrote about Anthropic being ordered to suspend access to both Fable 5 and Mythos 5 after the US government raised national security concerns. The latest twist is that Mythos 5 is now being allowed back, but only for a small group of approved US cyber defence organisations and critical infrastructure providers. Fable 5, the public model that triggered the original concern, remains restricted.
If the issue was simply that Fable 5 could be jailbroken, you would expect the public model to remain restricted while Anthropic fixed the problem. But the fact Mythos 5 is being allowed back for specific cyber defence users suggests the government also sees value in the capability it is restricting.
So we are more or less back in the holding pattern Anthropic was in before Fable 5 launched. Mythos can be used by a small group of trusted organisations, Fable is still stuck, and everyone else is waiting to see what the US government decides next.
There are also reports that Fable 5 could be allowed back soon, which makes the whole thing feel even more improvised. If that happens, then the big outcome of this episode may not be a permanent ban, but a new precedent: frontier model access can be interrupted by the US government after launch, negotiated behind closed doors, and then restored in stages.
There are clear similarities with the OpenAI story above. GPT-5.6 is being delayed into a limited preview, with customer access approved case by case. Mythos 5 is now back under a similar kind of carve-out. So this looks like the US government is starting to put itself between frontier AI companies and their customers.
Maybe that is inevitable. If these models really can help find weaknesses in critical systems, then governments are going to want some control over who can use them. But this still feels like regulation being invented after the fact, rather than a clear framework that companies, customers and other countries can plan around.
The Verge | TechCrunch | Gizmodo
Web 4.0
AI Ethics News
Nvidia wants to cut data center water use, but that’s not the same as fixing AI’s water problem
Google DeepMind bets $75M on AI’s future in Hollywood with A24 deal
UK’s top data and AI regulator quits after ‘inappropriate’ humour
Long Reads
Stratechery - Anthropic’s Safety Superpower
“The future is already here, it’s just not evenly distributed.“
William Gibson