AIHOY and happy weekend, AInauts!
Maybe you missed a few AI news, tools, and practical ideas this week. Or maybe you have just joined us. No problem.
Here are our highlights from the past week:
And at the end, you get the most important Quick News, so you can catch the relevant developments in one pass. Ready? Let's go.
The technical setup is rarely the bottleneck anymore. Creating skills, testing MCPs, starting a browser agent: all of that works surprisingly well now. The real bend in the road comes afterward, with one plain question: what should the agent actually do for you each morning?
In our AI Employee Bootcamp, we see exactly this gap. That is why we give you three prompts: find the best agent jobs from your real work, turn them into a clean briefing, and set up the agent directly on your machine. Plus we clear up the endless confusion around skills, MCPs, and agents.
Our advice: build one boring agent this weekend. One you can honestly evaluate after seven days. Did it save you 20 minutes every morning? That is enough for a start.
For three days, Claude Fable 5 was the model that made everything else look old: huge context, long autonomous tasks, finally a model that did not need hand-holding after every second sentence. Then the U.S. government pulled the plug.
On June 12, the Trump administration demanded that access to Fable 5 and Mythos 5 be blocked for all foreign nationals, even Anthropic employees in the U.S. There was no clean way to enforce that, so Anthropic switched both models off for everyone. The trigger was reportedly a jailbreak.
The lesson for us is uncomfortable: the strongest tool in your stack can disappear tomorrow, and you cannot stop it. A kill switch you do not control. That is exactly why local and open models are no longer hobby gear. They are basic infrastructure.
While everyone is mourning Fable, OpenRouter is making a cheeky argument: maybe one single model does not have to do everything anymore.
The new OpenRouter Fusion is basically a small model panel. You ask a complex question, several models work on it in parallel, a judge model compares the answers, and the system builds the final result. That is closer to good teamwork than classic routing.
For short prompts, Fusion is overkill. But anywhere a single answer is dangerously convenient, such as research, strategy, or comparisons, it gets interesting.
After the Fable shock, the same question kept coming up: how do I make myself less dependent? We sorted our answer into three stages.
Stage 1 is familiar from us: Files over Tools. Your knowledge lives in folders, not locked inside one app, so you can move from Claude to Codex to Antigravity without losing your work. Stage 2 makes you independent from one subscription: an OpenRouter key gives you access to hundreds of models from more than 75 providers. Stage 3 is the radical version: local models that run on your own machine.
Independence does not mean replacing everything with free tools. It means never being tied to a single provider again. Usually you give up a bit of frontier-model performance. For many use cases, that trade-off is completely manageable.
Chatting with a local model is nice. But many of us now work differently: the AI reads and writes in our folders and works through tasks step by step. Cowork-style work, basically.
That now exists in an open version too. OpenWork describes itself as an open-source alternative to Claude Cowork, a desktop app for Mac, Windows, and Linux. It connects to your local folders, handles multi-step tasks, supports skills, templates, and permissions, and works with your own OpenRouter key.
We tested it quite a bit and find it very interesting. Your folder stays the source of truth, and you pick the model from a dropdown. That way of working belongs to you. No subscription can cancel it for you.
To close, here is a small hack that connects directly to last week's loops and prompt tips. Meta-prompting used to be a big topic in our trainings: prompts that write better prompts for us.
Now we can lift that idea into the agent world. Instead of typing a clean briefing yourself, you let the AI write its own /goal. You give it the task, context, and boundaries, and it turns that into a main goal, success criteria, and stop rules. You only approve it.
That means you delegate the work and the briefing. And honestly, the AI often knows its own mechanics better than we do. This gets every task started cleanly, without you writing a full brief first.
AI News Quickie: the HAI-lights from the industry
AI never sleeps. This week was all about record deals and the biggest IPO ever, agents that suddenly pay and work on their own, and governments trying to slow AI down. Here are the most important stories.
🏛️ Anthropic, Claude, and the Fable aftermath
Fable 5 is still offline, but Anthropic sounds confident it will return soon. The White House reportedly feared Chinese access and gave Anthropic 90 minutes to shut it down.
More than 80 security leaders from Adobe, Zoom, Nvidia, and others are asking in an open letter for the restriction to be lifted. Some had warned in April how dangerous Mythos could be.
Anthropic is opening an office in Seoul and has signed a memorandum of understanding with Korea's science ministry.
Through an alliance with Tata Consulting Services, Anthropic is putting Claude in the hands of 50,000 consultants and moving into regulated industries. In other words: where the money is.
OpenAI under constant fire
OpenAI is making Shopping Ads in ChatGPT almost automatic: merchants connect their catalogs, and the system builds the ads.
Smart move: with Deployment Simulation, OpenAI replays old conversations through a new model before each release to predict bad behavior.
With LifeSciBench, OpenAI tests how well AI can handle real life-science research: conflicting data, incomplete evidence, and experiment design.
Companies now get Spend Controls in ChatGPT Enterprise: credit limits per team and person. Usage finally becomes visible.
OpenAI is buying cloud company Ona, formerly Gitpod, so Codex can run long coding jobs in the cloud even when your laptop is closed.
After the confidential S-1, OpenAI is now reportedly heading toward a Q4 IPO with Goldman Sachs and Morgan Stanley.
GPT-5.2 story: all chats are automatically moving to GPT-5.5.
🖥️ Google between launches and cleanup
With the Open Knowledge Format, Google is making the LLM wiki pattern official: knowledge as Markdown instead of proprietary catalogs. Files over Tools says hello.
The Antigravity CLI now replaces the old Gemini CLI. For many teams, that is a forced migration that breaks existing automations first.
DeepMind will treat its own agents like potential insider threats: rights are granted step by step, and monitoring flags suspicious behavior. Trust is good, sandboxing is better.
The June Pixel Drop brings new features such as text-to-video with Gemini Omni, Screen Reactions, and live speech translation to phones. AI keeps moving into the pocket.
Google is suing a Chinese AI fraud network that used Gemini to build scam pages and defraud hundreds of thousands of victims. AI helps both the attack and the defense.
Google's AI Overviews are telling users that horror-fiction creatures are real.
Agents learn to pay, work, and remember
Coinbase and AWS now let agents pay with crypto: through x402, a CloudFront provider can invoice the AI directly.
Visa follows and lets agents in ChatGPT handle real purchases, including spending limits and tokenization. Two payment rails, one goal: your agent pulls out the card.
Microsoft Copilot Cowork is live globally, and it runs on Claude Opus 4.8.
OpenAI is launching a Partner Network with 150 million dollars and Accenture, BCG, and McKinsey. 300,000 consultants are supposed to carry AI into companies.
Bedrock gets AgentCore Web Search and a managed knowledge base. AWS is building the workbench for enterprise agents.
Perplexity gives its agent a learning memory: Brain builds a context graph overnight from completed work and remembers what worked.
Sakana AI introduces Marlin, a virtual CSO: an agent that researches for up to eight hours and produces 100-page reports plus slides.
The deals are getting wilder
SpaceX, now merged with xAI, is swallowing coding star Cursor for 60 billion dollars in stock. Elon is filling the gaps in his empire.
Just days earlier, SpaceX itself went public: the largest IPO ever raised 75 billion dollars, and the stock jumped 19 percent on day one.
Salesforce is buying support platform Fin, formerly Intercom, for 3.6 billion dollars.
Snap is spinning out its expensive gen-AI video team into a separate company called Dotmo. Too expensive in-house, too valuable to throw away.
Mistral, Europe's answer to U.S. dominance, is negotiating 3 billion euros in fresh capital and aiming for a 20 billion valuation.
Perplexity is raising 200 million dollars to keep scaling its AI browser Comet.
🌍 Power, money, and infrastructure
Where is Germany's AI capital? The new AI Location Index 2026 compares 37 cities: Berlin leads clearly with 84.7 points ahead of Munich. Freiburg has the highest job density, and AI Engineer is the most sought-after profile.
At the G7 summit in Evian, Altman, Amodei, Hassabis, and others sat together in front of heads of state for the first time. Topic: AI sovereignty and fear of the kill switch. After the Fable drama, no surprise.
China is planning a 295 billion dollar national AI data-center grid, 80 percent powered by domestic chips. Nvidia is deliberately kept outside.
Oracle reports an AI order backlog of 638 billion dollars, up 363 percent. The stock still fell: that much backlog has to be financed and built first.
GitHub is routing traffic to AWS because AI coding agents are overrunning the infrastructure: 275 million commits per week and availability temporarily below 89 percent.
🛡️ Politics, society, and law
The White House is making its AI rules in real time. The Fable shutdown was the preview, and a clear line is still missing.
Amazon employees say they are being pushed out because they supported data-center limits.
A new Pew study shows that two thirds of Americans think AI is moving too fast, and 40 percent expect more harm than benefit. Even younger people are skeptical while still using chatbots heavily.
Actors' union SAG-AFTRA has ratified a four-year contract with the studios, including strict rules against digital doubles. More than 90 percent voted yes.
In the RIAA case against Suno, discovery reportedly shows millions of copied songs. A fair-use decision is expected in July.
Bitter punchline: the world's leading deepfake expert no longer trusts his own eyes. Media literacy is becoming a survival skill.
🎬 Tools, hardware, and the creative world
xAI's Grok Imagine Video 1.5 jumps to the top of the video arena, with sound and at a fraction of Sora's price. The AI video price war is on.
Adobe is putting its Firefly assistant into Premiere, Illustrator, and InDesign too: sort clips, clean up layers, and build brand kits by instruction.
Meta is rebuilding Facebook around AI Mode: instead of scrolling, you ask in normal language and get an answer from public posts, groups, and Reels.
Snap is selling its AR glasses Specs to everyone for a hefty 2,195 dollars.
Xreal joins in too: the Aura glasses run on Android XR with deep Gemini integration and weigh under 95 grams.
Epic is building generative AI directly into Unreal Engine, with Claude, Gemini, and Codex via MCP.
Pinterest launches the experimental shopping app Ask Pinterest plus AI ad tools and its own MCP. Inspiration turns into agentic shopping.
At Cannes Lions, NVIDIA and partners are reshaping advertising and marketing with AI.
Done. But no need to be sad. The AInauten will be back soon with fresh fuel for you.
See you Monday with a new round of news, hacks, and insights.
Your AInauten, Fabian & Reto
P.S. Follow us on the socials: X, LinkedIn, Facebook, Instagram, YouTube, TikTok.
