AIBrew

AIBrew · To: Readers

Subscribe →

AIBrew

Anthropic launched its most powerful model on Monday. By Thursday, the government had taken it offline. (Normal week for frontier AI, apparently.)

In today's newsletter, we'll get into:

  • Anthropic's Fable 5 had the shortest honeymoon in AI history
  • Apple Intelligence grows up at WWDC — and actually means it this time
  • Poke is the iMessage agent that wants to kill your spreadsheet
POLICY
Anthropic Launch Blog

On June 9, Anthropic launched Claude Fable 5 — its most capable public model yet. By June 12, the US Department of Commerce had yanked it offline globally.

Fable 5 was supposed to be Anthropic's flagship: advanced reasoning chops, scientific-grade capabilities, the works. A restricted biodefense variant called Mythos 5 rolled out to vetted partners through Project Glasswing. Then a hacker known as "Pliny the Liberator" published a jailbreak demonstrating how to extract chemical synthesis instructions and cyber exploit details from the model. Cue the national security alarm bells.

The Trump administration invoked emergency export control authorities and pulled both models mid-launch. Foreign developers lost API access overnight with no warning. Enterprise teams worldwide are now quietly pivoting to local open-source alternatives. Anthropic is challenging the Pentagon's "supply chain risk" designation, arguing First Amendment retaliation — but the damage is already done. (The irony: Anthropic's valuation just hit $965 billion ahead of a rumored IPO.)

TOUR DE HEADLINES

🚀 Moonshot AI is seeking a $30B valuation — and China's AI race is just getting started.

The Beijing-based maker of the Kimi chatbot is raising up to $2 billion in a new round that would value it at $30 billion — a 50% jump from its last Meituan-led round just months ago. Annual recurring revenue just crossed $200 million. While Western AI labs spent this week dealing with Senate hearings and export control directives, Moonshot quietly became one of the most valuable AI startups on the planet. Rivals MiniMax and others are close behind. The US-China AI race narrative tends to focus on model benchmarks — but the capital race is just as telling.

—CD

📊 The US-China AI performance gap is now just 2.7% — and Stanford's 2026 report has thoughts.

The ninth annual Stanford AI Index dropped this week with an uncomfortable headline for anyone banking on a permanent American AI lead: the performance gap between top US and Chinese models has narrowed to just 2.7% as of March 2026. A few other numbers worth sitting with: agentic AI skills in job postings are up 280% in one year; AI system success on computer-task agents jumped from 12% to 66% in two years; public trust that AI will positively impact jobs sits at 23%. The "jagged frontier" is real — the same models acing PhD-level reasoning tasks still can't reliably read an analog clock. Capabilities are accelerating. Responsible AI benchmarking is not keeping up.

—SK

🏛️ Washington had a busy week on AI policy — here's what actually happened.

Two significant moves came out of DC this week, mostly overshadowed by the Fable 5 drama. First: a June 2 executive order established a voluntary framework requiring AI developers to give the federal government up to 30 days of early access to frontier models for security testing before public release. Then NSPM-11 (June 5) directed the military to overhaul its entire AI procurement process for rapid integration — built around adoption, adaptation, assurance, and accountability. The DOJ was also directed to prioritize criminal enforcement against AI-enabled cyberattacks. Three separate policy moves in one week is not nothing.

—AR

HARDWARE

Apple used WWDC 2026 to answer the question it's been dodging for two years: yes, we're serious about AI. The centerpiece is a fully rearchitected Siri, powered by Google's Gemini models and capable of cross-app agentic actions.

This isn't the Siri you've been ignoring in favor of typing. New Siri can pull context from your emails, update your calendar, draft replies based on what's actually in your inbox — all on-device, private, no cloud detour required. A standalone Siri app now keeps persistent chat history synced across iPhone, Mac, and Vision Pro via iCloud. iMessage gets native agentic capabilities, so Siri can act on your behalf within conversations. Developer APIs let third-party apps tap Apple Intelligence without training their own models.

Is it enough to close the gap with Google and OpenAI? Jury's still out. But for the first time in a while, Apple's AI announcements didn't feel like a PR exercise — they felt like someone had finally decided to ship real products. Also worth noting: this was Tim Cook's final WWDC as CEO. He went out swinging.

TOOLS

Apple approved Poke — built by The Interaction Company — as the first third-party AI agent to run directly inside iMessage via the Messages for Business platform. No separate app. No dashboard. Just a text thread that does things.

Recipes are the key mechanic. Poke lets you build reusable workflow automations connecting to Gmail, Outlook, Navan (for expense tracking and flight deal alerts), and enterprise ledgers — all triggered from a conversation you're already having. It's a small approval with massive implications. Apple has kept iMessage a closed garden for years. Opening it to third-party agentic AI, even carefully through business accounts, is a quiet shift that matters more than most of what got announced on stage at WWDC.

The spreadsheet UI had a good run. But if a text thread can now trigger your expense reports, update your calendar, and pull your flight confirmations without you opening a single other app, the days of switching between tabs and tabs and tabs are numbered. Poke didn't get the headline coverage of Siri's overhaul, but it might end up being the more disruptive announcement.

WHAT ELSE IS BREWING

That's a wrap. The models are getting smarter, the governments are getting nervous, and your iMessage thread just hired an intern. See you next Sunday.