TL;DR: OpenAI is rolling out its upcoming GPT-5.6 model in phases—because the US government wants to vet it to ensure it’s safe, according to a report from the Information yesterday. It’s the government’s latest move to scrutinize the most capable models—and could be a preview of a slower era of AI that frustrates users and labs alike.

What happened: OpenAI is getting micromanaged. The federal government will reportedly be “approving access customer by customer” while GPT-5.6 remains in a preview limited to a small circle of partners, per recent comments CEO Sam Altman made in a company memo, which was reviewed by the Information.

It’s a far cry from the hands-off approach President Donald Trump took at the start of his second term—and it makes that “voluntary” review process from his AI oversight executive order look a little less optional.

OpenAI isn’t thrilled: “We’ve made clear to the US government that this is not our preferred long term model,” Altman reportedly wrote in the memo. And it’s not the only lab being “voluntold” what to do. Meta has faced pressure to submit its models for review, while Anthropic yanked Fable 5 and Mythos 5 two weeks ago under a Trump admin order to cut off access for all foreign nationals.

Move slow; don’t break things: The government is trying to pump the brakes on new model releases at a moment when AI companies are competing fiercely for customers—especially as they become more cost-conscious—and many are preparing for IPOs (though OpenAI’s could get postponed ‘til 2027.) And investors aren’t likely to be happy that the government can pull or stall these firms’ flagship products at will.

AI, interrupted: The slowdowns and disruptions aren’t just frustrating for labs, but for anyone who uses their AI—like the roughly 1 billion monthly active users ChatGPT now has. Meanwhile, the NSA was in the middle of testing Mythos when it got shut out along with everyone else. An AI legal startup even sued the government recently, claiming the loss of Fable did “immediate, irreparable, and existential” harm to its business.

Bottom line: AI companies have spent years warning that their increasingly advanced models could be dangerous. Now the government’s taking them at their word and tying releases up in red tape. —WK

Your new Slack coworker is Claude

Earlier this week, Anthropic launched Claude Tag, which lets you tag Claude in a Slack channel and hand off a task like pulling data, working through tickets, or digging up that one link someone posted three months ago (it then reports back in a thread).

Setup is admin-controlled, so your admin will decide which tools and data it can touch and pick which channels it lives in. Its memory stays scoped to each channel, so the Claude helping your sales team apparently won’t spill all the goss to the engineering department.

When added to a channel, Claude searches back through messages to catalog everything, then remembers what’s been said so you don’t have to re-explain context every time. Here are a few things I’ve picked up while testing it:

• Just like a regular chatbot, your wording matters when you search. I asked Claude to find a link a colleague had sent “about tokens,” and it surfaced the wrong post. That’s because the original message linked to something titled “why you keep hitting your usage limit” and never used the word tokens. Claude acknowledged the error and said: “I match on the words actually in the message, so the closer the search terms are to the post’s phrasing, the better.”
• You can get it to roast your coworkers. One colleague told Claude to “keep an eye on all posts in this channel and just roast everyone pedantically for everything they post.” It lasted about three posts before we tapped out—but it’s good to know it’ll run whatever (ridiculous) standing instruction you give it.
• It’s usage-based, and that adds up. Every task costs tokens, which one wary colleague figured could “end up costing a secret fortune.” Admins can cap monthly spend org-wide and per channel and see a log of every task and who ran it.

Tag is now in beta on Enterprise and Team plans. Eligible orgs get free launch credits to play with (the exact amount seems to depend on your plan), so it’s worth experimenting at least until those run out—then assess whether it’s worth keeping the meter running. —SM

For weeks now, Americans and Europeans have been locked in a bitter feud. The topic of dispute: how to deal with the summer heat. On one side, Americans are baffled as to why Europeans won’t just install air conditioning. On the other, Europeans shoot back that AC is expensive and doesn’t help when you’re not driving around everywhere.

The cooling gap across the Atlantic is, in fact, yawning—most US households have AC while under a quarter in Europe do, per energy sector estimates. The reasons are complicated, but boil down to a mix of electricity being more expensive, older buildings needing to be retrofitted, and a healthy dose of the belief that AC is a wasteful indulgence.

This thermostat standoff has led to a lot of mocking about “anti-AC cope,” increasingly off-topic roasts about each other’s weather and buildings, and appeals to the moral high ground. One person even dragged Claude into the argument (the chatbot is firmly in the Americans-are-right camp), and even Elon Musk weighed in. One thing’s for sure: Getting embroiled in a (ahem) heated debate with a whole other continent probably isn’t going to cool you off. —WK

Chaos Brewing Meter: /5