The Government Slowed Down GPT-5.6, Claude Moved Into Slack, and a Model That Changes Everything Just Showed Up

By Ivana Tilca · July 2, 2026 · 5 min read

The Week AI Stopped Moving Fast and Started Moving Carefully By Ivana Tilca · July 2, 2026 · 7 min read

The government slowed down GPT-5.6, Claude moved into Slack as a full teammate, and a model that doesn't build intelligence but orchestrates it quietly matched the frontier. Three stories, one theme: the era of "surprise us every Wednesday" is ending, and something more deliberate is taking its place.

Some weeks in AI are loud because a hundred things happen. This wasn't one of those. This week was loud because three things happened, and each of them says something uncomfortable and fascinating about where all of this is going.

I want to walk you through them the way I'd explain them to a friend over coffee — not just what happened, but why I think they belong in the same conversation. Because they do. Let's start with the launch that genuinely made me stop scrolling.

Sakana Fugu: the model that doesn't compete, it delegates

On June 22, Sakana AI released Sakana Fugu, and I need you to resist the urge to file it under "another new model." It isn't one.

Most models you know — GPT, Claude, Gemini — try to be the smartest brain in the room. Fugu takes the opposite bet. It's an orchestration model: instead of answering your prompt itself, it decides which model is best for the job, hands the task off, and sometimes coordinates several models at once before stitching their work into one answer. From your side, you call a single API. Behind the curtain, a whole team of experts is arguing it out.

Here's the detail that made me sit up. Sakana is explicitly pitching this as insurance against a very 2026 problem: single-vendor dependency. If one provider goes down, gets restricted, or disappears behind an export control overnight, Fugu just reroutes around it and your work never stops. Hold onto that idea — it's going to matter again in about three paragraphs.

It comes in two flavors:

Fugu — tuned for low latency. Fast, cheaper, and a sensible default for everyday work and coding tools.

Fugu Ultra — tuned for maximum quality on long, messy, multi-step problems. Slower and pricier, but it coordinates a deeper pool of experts when accuracy really matters.

And the benchmarks are the part nobody expected. VentureBeat's headline said the quiet thing out loud: no Claude Fable 5, no problem. Fugu Ultra stands shoulder-to-shoulder with frontier models like Anthropic's Fable 5 and Mythos Preview across serious coding, science, and reasoning tests — nudging ahead on some, tying on others, slipping only slightly on the hardest agentic ones. Think about that for a second: a router matched the models it's routing to. That's a real shift in what "state of the art" is even supposed to mean.

If you want to try it, you can spin up an account at console.sakana.ai with a pay-as-you-go plan and plug it into Codex. One honest warning, though — reviewers running Ultra at full tilt on ambitious projects reported bills climbing fast (one build-an-app-and-clone-a-site experiment ran up around $30). So start small before you hand it your dream project.

My take: as a reasoning and code engine, Fugu Ultra is impressive, and as a philosophy — resilience over raw scale — it might be the most quietly important idea of the week. It's not there on visual creativity yet. But that was never the point.

Claude moves into Slack — and stops being a chatbot

The Anthropic news is the one that genuinely excites me, because it changes the shape of working with AI, not just the score on a benchmark.

On June 23, Anthropic launched Claude Tag. In plain terms: you can now add Claude to Slack as if it were a member of your team. You give it access to specific channels and tools, and anyone can type @Claude — exactly like tagging a coworker — hand it a task, and go do something else while it works.

What makes this feel different from copy-pasting into a chat window comes down to four things Anthropic built in:

It's multiplayer. One Claude lives in the channel and everyone shares it, so your teammate can pick up a task where you left off.

It learns over time. The longer it lives in your channel, the more context it absorbs. You stop re-explaining your company from scratch every single time.

It can take initiative. With "ambient" mode on, it flags things it thinks you'll want to know and follows up on threads that went quiet.

It works asynchronously. Hand it a project and it can schedule its own steps and grind on it for hours.

And the number that stopped me: Anthropic says 65% of its product team's code is now written by their internal version of this tool, and it's spreading past engineering into support tickets, metrics, and bug hunts. When the company building the thing reorganizes its own work around it, that's worth noticing.

This is the part I keep thinking about. For years, "using AI" meant leaving your work to go talk to it — open a tab, paste context, copy the answer back. Claude Tag flips that. The AI comes to where the work already lives. That's arguably the biggest redesign of the interaction model since the standalone chatbot, and it points at a near future where the AI isn't an app you visit but a presence in your tools.

One catch: it's in beta for Slack Enterprise and Team customers only, running on Opus 4.8, and it replaces the old Claude in Slack app. Free-plan folks will have to wait.

The government put a speed bump in front of GPT-5.6

Now the story that made me raise an eyebrow — and the one that ties this whole week together.

According to TechCrunch, OpenAI is limiting the release of its new GPT-5.6 lineup — Sol (the flagship), Terra (the balanced one), and Luna (the fast, cheap one) — to a small group of trusted partners "whose participation has been shared with the government." This came at the request of the Trump administration, under a recent executive order asking frontier labs to submit their most advanced models for review up to 30 days before release.

And this isn't happening in a vacuum. Just weeks earlier, the same administration ordered Anthropic to pull foreign-national access to Fable 5, which led Anthropic to take the model down entirely before redeploying it globally on July 1. Two of the most powerful models on earth, both throttled by policy within a month of each other.

OpenAI complied — but made its feelings clear in a blog post: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

Here's the implication I think is worth saying out loud: we may be watching the end of the "surprise drop" era. That feeling of waking up on a random Wednesday to a shockingly powerful model nobody saw coming? If every frontier release now needs individual sign-off, that rhythm changes — and so does the pace of the entire ecosystem. Critics, including a former White House AI advisor, warn it edges toward a de facto licensing regime that could stall launches and hand time to competitors abroad.

And now — back to Fugu. Notice the loop closing? Sakana is selling resilience against exactly this: a world where access to any single model can shift overnight because of regulation, export controls, or policy. A week that hands you both the disease and someone's proposed vaccine is a rare thing. This was that week.

So I'll put the question to you: is government sign-off on frontier models a reasonable safety measure, or too much control over private technology? I genuinely want to know where you land.

Rapid fire — the rest worth your attention

ByteDance's Seedance 2.5 landed at its Beijing conference: 30-second clips at native 4K from a single prompt, up to 50 reference inputs (its predecessor took 12), 10-bit color, and audio synced natively to the visuals. Enterprise beta is live, public launch targeted for July. They skipped four version numbers just to make the "generational leap" point.

The Atlantic's AI training database. A searchable tool revealing which music and videos were used to train models like Suno and Udio — think 150 Blink-182 tracks, hundreds of individual YouTubers' videos. It's the kind of transparency we simply didn't have before, and it's going to fuel real fights over copyright and consent.

Meta Glasses. A refreshed line with 26 styles, new frames (Meta Adventurer, Meta Fury), a Kylie Jenner collaboration, and Meta's new Muse Spark AI from day one — starting at $299. If the original Ray-Ban look wasn't for you, there's more to choose from now.

What I'm taking away from this week

Here's the thread I can't unsee: for years, progress in AI meant bigger and faster — bigger models, faster drops, a new frontier every few weeks. This week, all three big stories pointed somewhere else. Fugu bets on coordination over scale. Claude Tag bets on fitting into your day instead of demanding your attention. And the government bet — for better or worse — on slowing the release valve.

Maybe that's the real headline. AI isn't just getting more powerful; it's starting to get more deliberate — in how it's built, how we use it, and who decides when we get it. Whether that deliberation protects us or just slows us down is the question we'll be living with for a while.

That's what I try to do here: drink from the firehose, filter out the hype, and hand you back only the parts that actually move the needle. Not every launch is a revolution. But some weeks quietly redraw the map — and this was one of them.

Which of the three hit you hardest: the orchestrator, the teammate, or the government stepping in? Tell me in the comments. I read them, and this one I actually want to argue about.