Çelik here

On Friday the most powerful AI model on the planet stopped working, not because it broke, but because Washington said so. In the same four days a Chinese lab put a model of nearly the same caliber on the internet for anyone to keep forever, Salesforce paid $3.6 billion to own an agent instead of renting one, and the OECD quietly graded which Balkan economies will have their own compute when the next model goes dark. One thread runs through all of it: this week separated the AI you access from the AI you own.

If you build on a US frontier API, the Anthropic recall is the wake-up call for your whole stack. If you operate in Kosovo, Macedonia, or Albania, the OECD scorecard says where regional compute is heading and who gets left renting.

And the MiniMax release at the bottom is the cleanest answer to the question the week keeps asking: what do you do when access can be revoked?

On June 12 the White House ordered Anthropic to cut off Fable 5 and Mythos 5 for every foreign national, inside or outside the United States, citing export-control authority after officials learned of a jailbreak that could expose cyber and vulnerability-discovery tricks. Compliance was all or nothing, so Anthropic disabled both models worldwide, for paying customers and its own foreign-national staff alike. The most capable models the company ships went dark in a day, by directive, not by outage.

The positional shift is that the model under your product is now a licence the US government can revoke. Anthropic is contesting the order, warning the standard would freeze every new frontier deployment in the industry, but the precedent is set: a model can be a strategic export one morning and unavailable the next. Anyone outside the US who wired a frontier API into their stack just found out they are a foreign national in someone else's export regime.

Inventory which of your products lean on a single US closed model with no fallback, and stand up a second provider you can switch to inside a day. Watch whether Washington and Anthropic settle this week, because that readout tells you whether this was a one-off or the new shape of access. The model is no longer just a vendor choice, it is a jurisdiction.

Busy executive. Packed calendar. An inbox that never empties.

Everyone is talking about AI agents. Every week brings another promise of an assistant that can do it all.

Catch is the real deal.

A smart, proactive AI admin agent focused solely on taking administrative work off your plate.

It schedules meetings, triages your inbox, drafts emails in your voice, resolves conflicts, sends follow-ups, and handles the countless small tasks that consume your day.

Available wherever you work — Gmail, Outlook, Slack, WhatsApp, and even over the phone.

No setup. No training. No learning curve.

Catch learns how you work, takes action when it's confident, and keeps things moving without constant supervision.

From swamped to sorted in seconds.

Get started with Catch and have your assistant ready before your next meeting.

Test Catch for free

The same week a US order pulled a frontier model from every foreigner, the OECD graded the Western Balkans on the exact capability that now decides who is exposed: access to compute, financing, and AI readiness. The scorecard, published June 9 to 11, lands as Skopje and Belgrade plug into Europe's own AI grid through new EU AI Factory Antennas, shared supercomputing nodes that give local startups and researchers sovereign access. Kosovo and Albania are not on that list.

For the region, the recall makes the divide concrete. A founder in Skopje with a line into European public compute has a fallback the day a US model disappears; a founder in Pristina or Tirana renting only foreign APIs does not. Sovereign compute stopped being a Brussels talking point and became the difference between a product that keeps running and one that stops.

If you build in Kosovo or Albania, treat compute access as a sourcing problem now, not a policy wish: map which EU programmes you can join through partners across the border, and price an open-weight fallback into your architecture this quarter. The economies inside the grid will compound the advantage, and the ones outside will keep renting on terms someone else sets.

On June 15 Salesforce agreed to buy Fin, the customer-service AI company formerly called Intercom, for about $3.6 billion, folding an agent that closes roughly three of four support tickets on its own into Agentforce. The same day, Arcade.dev raised $60 million to build the authorization layer that decides what a production agent is allowed to do. One deal buys the agent, the other sells the leash.

Read the two together against the recall. Incumbents are paying up to own the agent layer outright rather than rent it, precisely because rented capability turned out to be revocable, and capital is rotating toward the plumbing that governs agents in production. Even the week's other record points the same way: SpaceX closed its first full week up 43% from its IPO price and above $2 trillion, the market paying a premium for assets you control rather than access you borrow.

Look at your own stack and separate what you own from what you rent, then decide which rented pieces are load-bearing enough to bring in-house or wrap in a control layer of your own. The agent gold rush is entering its buying phase, and the companies with cash are choosing ownership over dependency.

While one frontier model vanished by government order, another the same month became impossible to recall. Chinese lab MiniMax released M3, the first open-weight model to pair frontier coding, a one-million-token context window, and native multimodality, with the weights and technical report posted for download around June 11 for anyone to run on their own hardware. On SWE-Bench Pro it scores 59.0%, edging GPT-5.5.

The contrast is the whole issue in two models. A closed model is a service the country that hosts it can switch off; an open-weight model on your own cluster cannot be revoked by anyone, because there is no access to cut. When a frontier-grade open model runs on private hardware, sovereign AI stops being a procurement slogan and becomes a download.

Pull M3 down and benchmark it against whatever closed model your product leans on, not to switch tomorrow but to confirm your fallback is real before you need it. The week keeps saying the same thing in different languages: the model you can host is the only one no government can take back.

Five tools to install or test this week. Marketing and sales were quiet, so dev gets two.

Productivity: Microsoft Copilot Groups turns the assistant into a shared room. Microsoft added Groups to Copilot, a shared space where up to 32 people work with the same assistant on a document or project, alongside a new avatar called Mico. It is live in the US and rolling to the UK and Canada next. Put your next planning session in a shared Copilot instead of pasting outputs around.

Sales: Zoom ZoomMate finishes the work a call starts. Zoom launched ZoomMate on June 1, an agent that pulls a Salesforce record mid-call, logs the summary, updates the opportunity, and drafts the follow-up without switching apps, at $20 per user a month. It also reaches into Jira, Slack, and ServiceNow. If your reps live in back-to-back calls, run one week of pipeline through it.

Cost: Anthropic put a meter on agentic coding. From June 15 Anthropic moved Agent SDK and claude -p usage off flat subscriptions and onto $20 to $200 monthly credits. If your unit economics quietly assumed unmetered agent runs, re-model them this week before the credit ceiling does it for you.

Dev: Gemini CLI shuts off June 18. Google retires the Gemini CLI hosted path for Pro, Ultra, and free accounts on June 18, replacing it with the Go-based Antigravity CLI that keeps your skills, hooks, and subagents. Migrate before the deadline rather than after your scripts stop returning answers.

Dev: Microsoft Foundry Hosted Agents give each agent its own sandbox. Microsoft introduced Hosted Agents in Foundry Agent Service, a managed runtime where every session gets dedicated compute, memory, and a durable filesystem, framework-agnostic across the Agent Framework, Copilot SDK, and LangGraph. Stand a production agent on it before you widen its autonomy.

Next edition soon, Çelik