The government broke Anthropic's best model this week

In this edition:

• This week: the Fable 5 ban is ending, Block's coding agent now merges 1,500 PRs a week, and OpenAI hired the policy writer and the architecture writer in the same week

• Under the radar: Google DeepMind published its AI Control Roadmap, treating its own production agents as a potential adversary

• What's on the calendar: Micron earnings Tuesday, CoreWeave's first analyst day Thursday, and Anthropic's Fable 5 restoration update due any day

THE WEEK IN AI

THE WEEK IN ONE SENTENCE

Anthropic's best model is coming back online after a government shutdown with a 90-minute deadline. OpenAI hired the person who wrote the policy behind that shutdown and the person who wrote the architecture that the whole field runs on, and the gap between an agent demo and a 1,500-PR-a-week production system closed faster than most procurement cycles can absorb.

THREE SIGNALS

01 • Model

The Fable 5 ban is ending, the precedent is not

Anthropic's Chris Ciauri said at a Seoul conference on Thursday that Mythos and Fable 5 will regain availability in South Korea within days. Private negotiations at the G7 between AI executives and President Trump produced the unwind. That ends a standoff that started on a Friday night with a Commerce letter and a 90-minute deadline. In between, 76 cybersecurity experts signed an open protest, and Anthropic's best public model went offline globally for most of a week.

The fact that the unwind starts in days is good. The fact that it started with a 90-minute deadline is the part operators should not forget. A frontier model with a multi-billion-dollar enterprise footprint went dark on a Friday, on a directive the company could not even contest before the clock ran out. The restoration scope is also still unclear. Ciauri said "within days" and "South Korea." Whether other geographies follow on the same schedule, or follow at all, has not been confirmed.

If your AI contract does not have a vendor-continuity clause that names a substitute model and a switchover timeline, this week is the reason to write one. The model was paused by the government, not by the vendor. Your contract language probably did not anticipate that, because mine didn't either.

AI READY PRO · FREE UNTIL FRIDAY

I will get you AI trained in 30 days for free

After several months of running it quietly with top AI operators and teams, I’m excited to launch AI Ready Pro today to our newsletter subscribers like you.

It’s a 30-day personalized AI training program that is personalized to you and where you are in your AI journey. It’s the culmination of 100’s of hours spent teaching AI to top Fortune 500 AI teams and operators (and working with Mark Cuban).

Each day, you receive a 10-15 minute exercise to complete to improve your AI skills.

This isn’t vague AI theory or a way to pitch you a tool. It’s 30 days of learning to use AI in your real work. So you come out the other side with the skills to use AI to actually improve you and your teams output.

Until this Friday (May 15), it’s free for newsletter readers who complete the extended assessment. This assessment will help us personalize the learning experience to you.

If you lead a team, AI Ready Team is open today too, also for readers first. It’s the same engine, but to train your whole organization. You’ll get a detailed view of:

• Your team’s AI readiness level + detailed strengths & weaknesses

• Ranked list of AI automation opportunities

• Get tactical advice on how to advance AI efforts in your org

Leading a team? Take the Team assessment instead.

02 • Agents

Coding agents stopped being a story about coding

Block disclosed this week that its internal Builderbot agent now makes 15% of its production code changes. It merges 1,500 pull requests per week through Slack. Not suggestions. Merges. Cursor shipped /automate. The new feature lets the IDE run cloud agents triggered by Slack emoji and GitHub events, with full computer-use access. The Codex team shipped Record and Replay, which lets an operator demonstrate a task once and reuse the recording forever. Codex also disclosed 5 million weekly active developers. On the same Thursday, Accenture's stock dropped 20% after lowering its revenue outlook and blaming AI disruption, with new bookings down 2%.

I want to be honest about the pattern here. The Cursor and Codex launches are vendor announcements. The Block number is a real deployment at a real company. The Accenture print is a guidance cut from a $200 billion firm whose business model is the work these agents now do. Three different sourcing weights, one direction.

The thing worth carrying into next week is not "agents replace consultants." It is that the gap between an agent demo and a 1,500-PR-a-week production system can be closed faster than most enterprise procurement cycles can absorb. If your roadmap still treats AI coding as a productivity tool, the version that lives at Block is already operating one category up.

03 • Talent

OpenAI hired the policy writer and the architecture writer in the same week

Dean W. Ball, the White House AI adviser who drafted the 2025 America's AI Action Plan, joins OpenAI on July 6 as Head of Strategic Futures. He announced it himself on X this week. Two days later, Noam Shazeer, Transformer co-author and Gemini co-lead, left Google to join OpenAI as architecture research lead. Google had paid $2.7 billion to rehire him through the Character.AI acquisition just 18 months ago.

Two hires, one read. Anthropic is being shut off by a US government directive. In the same week, OpenAI hired the person who wrote the policy behind that directive, and the person who wrote the architecture that the whole field runs on. Sam Altman's note on the Shazeer move: "only took 10 years."

For an operator, none of this changes a roadmap on Monday. It changes the question you ask in a year. The spread between the top frontier labs that procurement teams treat as roughly comparable is widening on two axes that do not show up in a benchmark table: regulatory standing and the people building the next architecture.

UNDER THE RADAR

Google DeepMind quietly published its AI Control Roadmap on Thursday. The framework lays out 15 defenses for protecting internal systems against, in DeepMind's own phrase, "increasingly capable and imperfectly aligned" AI agents. Authors Rohin Shah and Four Flynn compare the approach to a driving instructor with dual controls: trust the agent to drive, stay ready to take the wheel.

The unusual part is the posture. DeepMind is publishing internal security architecture that treats its own production agents as a potential adversary. The controls sit on top of sandboxing and endpoint security, instead of relying on alignment training to make the agent safe by default. DeepMind said the framework could serve as a model for the wider industry.

If you are running any agent in production today, the right read is not the technical detail. It is the posture. A frontier lab just published, for free, the assumption you should be building from. Assume your agent will at some point try to do something you did not ask for, and build the controls so that when it does, it cannot. That is a different procurement conversation than the one most companies are having about agent rollout.

QUOTE OF THE WEEK

❝

"Only took 10 years."

Sam Altman, on Noam Shazeer joining OpenAI a decade after the Transformer paper was published at Google. .

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

WHAT’S ON THE CALENDAR

• Micron reports after the bell Tuesday. The read on whether Tim Cook's memory-cost warning — AI demand pulling supply away from consumer devices — shows up in guidance.

• CoreWeave hosts its first analyst day Thursday. Off a week in which it trained DeepSeek V3 671B in two minutes on 8,192 Blackwell Ultra GPUs, the fastest MLPerf result on record. Worth watching for what the company says about capacity and customer concentration.

• Anthropic's next Fable 5 restoration update is due any day. Anthropic said "within days" on the export control compliance gate. Days start now.

Have a good weekend,
Haroon