1. Fable is still dark, and the “misunderstanding” looks less like one

A week after Commerce switched off Fable 5 and Mythos 5, the models are still gone, and Washington has not softened. Anthropic keeps calling it a misunderstanding. The backstory says otherwise: Trump pulled federal agencies off Anthropic in February over contract terms, the Pentagon branded it a “supply chain risk” in March, and Anthropic is fighting that in court.

This is not a safety story you can engineer around. It is a power story your workflow got caught inside.

2. The same fight moved up the stack, into your software

While the model layer made headlines, the platforms quietly bought the layer above it. Salesforce signed a definitive agreement for Contentful to give Agentforce a content engine it owns. Contentful is German, and it now falls under US law, which Drupal founder Dries Buytaert summed up as “European today, not European tomorrow.” The same control question from the Fable story, one layer down, inside the tools you already pay for.

The platforms are not selling you better agents. They are buying the ground your agents stand on.

3. Agents are failing everywhere, and intelligence is not the reason

The demos dazzle and the deployments stall. Gartner expects more than 40% of agentic AI projects to be canceled by 2027, blaming runaway costs, fuzzy value, and weak risk controls, not the models. Teams buy the tool and launch the pilot before anyone decides who owns the workflow when the agent ships the wrong thing. For marketing specifically, Gartner warns a third of companies will damage customer experience by deploying too early.

The bottleneck was never intelligence. It was governance, the unglamorous work nobody demos.

4. The shutdown is now your compliance problem, and the fix is a backbone you own

Fable’s blackout did not stay Anthropic’s problem. Lawyers are telling companies to map every system that could route a foreign national to an export-controlled model and stand ready to reroute. The lesson is the one this newsletter keeps arriving at. Open models you can self-host, like NVIDIA’s Nemotron line.

A model someone else can switch off is a convenience. A model you run yourself is infrastructure.

5. Watch This Week: the terms of Fable’s return

The White House has signaled it wants Anthropic to remediate so the control can lift, but the path back appears to run through citizenship verification, which could return Fable to US citizens first.

The question is no longer whether Fable returns. It is whether “who are you, and where” becomes a standing input to what AI you are allowed to use.

A model you rent, a stack you do not own, a government that can reach into both. The models are extraordinary, which was never the question. The question is what you build on. The human context you bring is the one input no acquisition and no letter from Washington can switch off.

Context, as usual, is required.

The Bench

• Google Cloud made Gemini the default model inside Workday’s Sana agent.

• UnitedHealth committed $3 billion to AI, with agents touching 26 million consumers.

• Cloudflare rolled out temporary accounts for AI agents.

• MIT Sloan and BCG: agentic AI hit 35% adoption in two years, most stuck in pilot purgatory.

The story this week is not that AI got regulated. It is that the question every institution has been circling for months, who actually controls the AI you depend on, stopped being theoretical. On Friday afternoon the US government answered it by reaching into a private company and switching off the two most capable models on the planet. Almost no one who relied on them got a vote.

Five signals that access, not capability, was the thing you should have been worried about all along.

1. The government switched off the best AI on earth, and you found out you were renting it

Three days after launch, Anthropic’s two most powerful models went dark for every customer in the world.

Anthropic released Claude Fable 5 on June 9, the first time it put its top “Mythos-class” capability in front of the public, wrapped in safeguards meant to block the riskiest uses. On Friday June 12, the company says it received a Commerce Department export-control directive at 5:21pm Eastern ordering it to suspend Fable 5 and Mythos 5 for any foreign national, inside or outside the country, including Anthropic’s own non-citizen employees. Axios reports the letter came from Commerce Secretary Howard Lutnick and that the administration acted after another company claimed it could jailbreak the model, having already tried and failed to get Anthropic to delay the release.

To comply, Anthropic disabled both models for all customers globally. Its other models still work. Open a Fable session now and it ends in an error, with new requests routed down to older models like Opus 4.8. This is the first time the US has aimed export controls at an AI model itself rather than the chips underneath it, and a license is now required to export, re-export, or even domestically transfer the models.

Anthropic’s position is that this is a misunderstanding. The company says the jailbreak it was shown was narrow and non-universal, amounting to asking the model to read a codebase and fix flaws, and that other publicly available models can already do the same. The government has not made its specific concern public. Both of those things can be true at once, which is the uncomfortable part.

The capability was never the fragile thing. Your access was. A model that someone else can switch off by letter on a Friday afternoon was never your infrastructure. It was a rental you mistook for a foundation.

2. The Copilot bills landed, and they are a preview of the real price of everything

The other way you lose a tool you do not control is slower, and it arrives as an invoice.

GitHub Copilot’s flat plans converted to token-metered “AI Credits” on June 1, and as the first metered bills land, developers are reporting sharp jumps and calling it the end of Copilot’s flat-rate era. I do not think Copilot’s pricing is the story. It is the first consumer-facing crack in a structure this newsletter has been pointing at for two editions: the enterprise token bills that burned through Uber’s annual AI budget in four months were not an anomaly, they were an early look at the real number.

This entire industry is floated on venture money, and most of us are paying prices that have little to do with what these tokens cost to run. Flat pricing was a customer-acquisition subsidy, not a business model.

The question is not whether prices rise. It is whether your workflow still makes sense when a token costs what a token actually costs.

3. “AI psychosis” names the gap between the demo and the work

The financial bill is not the only one coming due. There is a human one, and a Box executive gave it a name.

Box CEO Aaron Levie coined the term “AI psychosis” in a late-May post, describing how leaders stay so far from the actual work that they fall for the happy-path demo and miss the verification, integration, and judgment steps that turn a demo into something that holds up. He was describing executives cutting jobs on the strength of a demo, with tech layoffs in the first five months of 2026 running close to the total for all of last year.

The same blindness is in all of us, not just the corner office. We have been handed extraordinary capability at a subsidized price and quietly assumed it will always be there and always stay this cheap.

The happy path is a demo. The last mile is a job. The companies firing the second to buy more of the first are about to discover the model still needed the human it replaced.

4. The web’s value flows one direction, and creators are being asked to subsidize that too

The control problem runs all the way down to whether anyone can find your work.

AI crawlers take far more from the open web than they return, with Anthropic’s crawl-to-referral ratio running into the thousands to one and most operators returning a tiny fraction of what they take. The instinct is to block the crawlers, but a blanket block opts you out of both training and AI search at once, because training and retrieval bots are now separate user-agents and most people block the wrong one by accident. Meanwhile Google used new Search Central guidance this week to position its own documentation as the primary reference for SEO, AEO, and GEO, telling businesses to distrust third-party tools claiming its endorsement.

It is the same lesson as the Fable blackout, one layer down. Discoverability is becoming something the platforms grant, not something you earn.

If your entire reach depends on a system you do not control, you do not have reach. You have a permission slip.

5. Watch This Week: whether Fable comes back, and what the precedent costs everyone

An administration official told Axios access could be restored “in the next few weeks,” once the government’s security apparatus is “hardened.” So the immediate outage may be short. The precedent is not.

Export controls now apply to frontier models the way they have long applied to weapons and chips, and a federal letter can take the best tool on the market off the board overnight. VentureBeat’s read for enterprises is blunt: any workflow tied to a single closed-API provider just became an operational risk, not a convenience. Watch whether this stays a one-off or becomes a tool the government reaches for again, and whether serious teams start treating multi-model fallback and self-hosted options as a requirement rather than a hobby.

The story is not whether Fable returns. It is that “the government can turn it off” is now a line in every honest AI risk assessment. Build like the switch exists, because this week proved it does.

The arc of these last three weeks is hard to miss. First the bill came due. Then the institutions started naming AI as a problem of power rather than performance. This week, someone actually pulled the lever.

None of this is an argument against the technology. The models are extraordinary, and they let a small operator build things that were out of reach two years ago. It is an argument about foundations. The most powerful model in the world is a tool you rent. The backbone of your work should be made of things you can run and understand yourself, even when they are slower and less impressive than what just went dark. Less impressive that survives beats spectacular that disappears.

And there is the quieter risk underneath all of it. If everyone runs their thinking through the same handful of models, everyone starts to sound the same, and sameness is the exact ground the giants win on. The human context you bring is the one thing no letter from Washington and no pricing change can switch off. It was always the point.

Context, as usual, is required.

James

Bench

• Apple shipped a rebuilt Siri and expanded Apple Intelligence at WWDC across its device ecosystem, while the EU confirmed the new features will not reach European users.

• A GPT-5.6 “kindle-alpha” checkpoint surfaced in developer channels with no official release.

• SpaceX began trading on Nasdaq under SPCX on June 12. The S-1 that revealed Anthropic’s $1.25 billion per month compute bill is now a public company. Infrastructure is still the actual market.

Share

1. The Pope, the CMA, and the SEC Walk Into Your AI Strategy

Three institutions that do not typically move together on technology reached similar conclusions within the same two-week window.

On May 14, the UK Competition and Markets Authority opened a Strategic Market Status investigation into Microsoft’s business software ecosystem, covering Windows, Office, Teams, and Copilot across 15 million commercial users. The CMA’s explicit framing is notable: the regulator cited “a shift towards agentic AI in familiar workplace tools” as part of its rationale for moving now. This is not a legacy antitrust concern. It is a forward-looking one about who controls the AI layer that enterprise workers operate inside.

On May 25, Pope Leo XIV released Magnifica Humanitas, his first encyclical, a 42,000-word document addressed to 1.3 billion Catholics focused on “safeguarding the human person in the time of artificial intelligence.” Leo called for AI to be “disarmed” and directed toward the common good, warning against the concentration of computing power in the hands of a small number of private actors. Anthropic co-founder Chris Olah attended the Vatican presentation.

The SEC’s Division of Examinations named AI-related misrepresentation an explicit priority in its 2026 agenda, committing to review registrant disclosures for accuracy regarding AI capabilities and to assess whether firms have adequate controls supervising AI tools. Taken together with Colorado’s AI Act taking effect June 30 and California’s AB 2013 requiring generative AI developers to disclose training data, the regulatory envelope is tightening from multiple directions simultaneously.

When a competition authority, the Vatican, and securities regulators reach convergent conclusions in the same fortnight, the “we’re being responsible” era is over. Governance is no longer optional or aspirational. It is the product.

2. Microsoft Build and the New Enterprise Control Layer

At Microsoft Build on June 2 and 3, the company heavily emphasized agentic workflows and cost-controlled orchestration inside Azure and GitHub Copilot. The timing was pointed. Microsoft walked into its developer conference having just canceled internal Claude Code licenses for thousands of engineers and redirected them to Copilot CLI by June 30.

Satya Nadella’s positioning was deliberate: Microsoft is the company that can deliver AI at scale without the uncontrolled token bills that burned through Uber’s entire 2026 AI budget in four months. The argument was less about benchmark capability and more about financial model design, enterprise integration, and governance architecture that a CFO can actually defend.

Early signals from the enterprise market suggest buyers are responding more to cost predictability than to raw model performance. Gartner placed generative AI in the trough of disillusionment throughout 2026, predicting enterprises will defer 25 percent of planned AI spending into 2027 as financial rigor slows production deployments.

The next phase of enterprise AI adoption will be won on financial model design, not benchmark scores.

3. Apple’s WWDC AI Moves

Apple opened WWDC today with anticipated announcements on Siri upgrades and deeper Gemini integration across its device ecosystem. With over 2 billion active devices, Apple’s on-device AI moves carry structural weight that enterprise-focused announcements often do not.

The question worth watching is not what Apple announces but how fast it moves from announcement to adoption at scale. Apple’s distribution advantage is unmatched. If Siri becomes genuinely useful for knowledge work tasks, the consumer AI experience may start shaping expectations inside enterprise organizations rather than the reverse.

4. Anthropic’s Compute Reality Check

SpaceX’s S-1 filing, amended June 1, disclosed that Anthropic is paying $1.25 billion per month through May 2029 for exclusive access to Colossus 1 and Colossus 2 in Memphis, covering approximately 325,000 Nvidia GPUs. Google separately disclosed a $920 million per month deal with SpaceX on June 5, covering a separate pool of roughly 110,000 GPUs through June 2029.

For context: Anthropic’s monthly compute commitment to SpaceX is nearly equal to the entire quarterly revenue of Starlink’s connectivity segment, which serves 10.3 million subscribers across 164 countries.

This is what frontier model development actually costs. The organizations debating which model to use in their AI strategy are operating several abstraction layers above the real competition, which is about who can afford to train the next generation of models at all. The moat in frontier AI is not the algorithm. It is the ability to write a $1.25 billion monthly check.

5. Watch: EU AI Act Implementation

The EU AI Act continues its phased rollout toward full applicability in August 2026. Several member states are actively pushing for targeted delays and simplification measures ahead of the deadline, particularly around the high-risk AI system classifications that will affect healthcare, hiring, and financial services applications.

The regulatory environment in Europe is not getting softer. It is getting more precise. Organizations with EU exposure that have been waiting for clarity before building compliance programs are now running out of runway. Modelling multiple scenarios is no longer a planning exercise. It is a legal one.

The institutions that matter are no longer asking whether AI works. They are asking who controls it, who pays for it, and who is responsible when it doesn’t.

That is a different conversation than the one most AI strategies were written for. The organizations that understood this six months ago are already rewriting their governance frameworks. The ones that are still in pilot mode are about to find out that the external environment has moved on without them.

Context, as usual, is required.

James

Subscribe now

Bench

• Alphabet reportedly raised $80 billion to extend its compute buildout, continuing the infrastructure arms race the SpaceX deals make visible.

• OpenAI reportedly landed a major cloud distribution deal with AWS, adding distribution muscle ahead of its anticipated IPO.

• Webflow announced a second 8 percent workforce reduction explicitly tied to agentic AI, continuing the pattern from ClickUp last week.

On May 14, Microsoft began revoking internal Claude Code licenses across its Experiences and Devices division, the team that builds Windows, Microsoft 365, Outlook, Teams, and Surface. The deadline is June 30. Engineers are being redirected to GitHub Copilot CLI, a cheaper tool Microsoft already owns.

The reason is not that Claude Code failed. The reason is that it worked too well. Six months ago, Microsoft rolled it out to thousands of engineers and encouraged everyone to reshape their workflows. Usage exploded. The token bills followed.

Uber confirms the pattern from a different angle. Uber’s CTO told The Information that the company burned through its entire planned 2026 AI coding budget in four months. Claude Code adoption had climbed from 32 percent to 84 percent of its roughly 5,000-engineer organization by March. Individual engineers were spending between $500 and $2,000 per month on tokens.

The trap is structural, not incidental. Flat seat licensing obscured actual token consumption until the invoice arrived. Most enterprise procurement and finance teams have no frameworks to forecast or cap usage-based token spend. They bought access like they buy software licenses. AI doesn’t bill like software.

If your organization is deploying frontier AI tools at team scale and tracking success by adoption rate, you are likely three months from a CFO conversation you are not ready for. The measurement system needs to change before the budget cycle does.

2. Two CEOs walked back the jobs apocalypse the same week OpenAI filed to go public

On May 26, Sam Altman told a Commonwealth Bank of Australia conference that he had been “pretty wrong” about AI’s impact on jobs, saying he expected more displacement of entry-level white-collar roles than has actually happened. Dario Amodei made an almost identical reversal the same week, prompting Fortune to label it a coordinated industry-wide walk-back.

The timing requires acknowledgment. OpenAI reportedly filed IPO paperwork confidentially in May 2026, targeting a $1 trillion valuation. A calmer narrative on job displacement is substantially better for a public offering than a jobs apocalypse framing.

The data offers a messier picture than either position suggests. The Yale Budget Lab found no significant change in unemployment for workers in high-AI-exposure jobs through March 2026. At the same time, 113,000 tech workers have been laid off across 179 companies since January 1, a pace 33 percent higher than the same period in 2025. Meta fired 8,000 employees on May 20, six days before Altman’s speech, while announcing $125 billion in AI infrastructure spending.

The honest read is not that the apocalypse was wrong. It is that the org chart moves slower than the technology does. The displacement exists. The timelines were off.

3. Practical: Opus 4.8 shipped and the headline is about honesty, not capability

Anthropic released Claude Opus 4.8 on May 28. Anthropic described the model as “a more effective collaborator,” which is a restrained framing for a flagship release.

The meaningful change is behavioral. Early testers report Opus 4.8 is more likely to flag uncertainties and less likely to make unsupported claims. A new dynamic workflows feature in Claude Code lets a single agent plan, distribute work across parallel subagents, and merge the result inside one session. Fast mode now runs at 2.5x speed and is three times cheaper than previous models.

This is the release Anthropic built for production deployments rather than demos. A model that bluffs less is worth more in agentic workflows than a model with better benchmark scores that confidently produces broken output. If you are evaluating whether to move agentic workloads from 4.7 to 4.8, the question to test is not whether it scores higher. It is whether it fails more transparently.

4. Hype Check: The model release cadence is becoming noise

GPT-5.5 launched April 23. Opus 4.8 launched May 28, just 41 days after Opus 4.7. Polymarket currently prices a 47 percent chance that GPT-5.6 arrives June 8 to 14. That puts three major frontier model releases inside a roughly seven-week window, with a fourth potentially landing before this edition is two weeks old.

The cadence has crossed a threshold. When model versions ship faster than most organizations complete an evaluation cycle, the version number stops being a useful signal. Gartner placed generative AI firmly in the trough of disillusionment in its January forecast, predicting enterprises will defer 25 percent of planned AI spending into 2027 as financial rigor slows production deployments.

The organizations chasing each release are not gaining a compounding advantage. They are incurring switching costs on infrastructure that will be obsolete before it is fully deployed. The asymmetric move is to stabilize on a model version that is good enough, build the orchestration and governance layer properly, and let the benchmark competition play out without you.

5. Watch This Week: Microsoft Build opens June 2

Microsoft Build runs June 2 to 3 in San Francisco, condensed to two days with a sharp focus on AI agents and enterprise developer trust. The expected announcements include Azure AI Foundry updates, GitHub Copilot multi-agent orchestration, and Satya Nadella’s public framing for Copilot as an agent-first platform.

Nadella is walking into a developer conference having just canceled a rival’s product internally and redirected his engineers to Microsoft’s own tool. The positioning argument he makes for Copilot this week is not purely technical. It is a defense of the enterprise cost model Microsoft controls versus the token-based billing model it does not.

Watch whether Build produces concrete pricing commitments or governance frameworks for agentic deployments. Announcements without unit economics are the pattern that got Uber and Microsoft into this problem in the first place.

The enterprise AI story in May 2026 is not that the tools stopped working. The tools worked fine. The problem is that nobody designed the financial model around what happens when they work at scale.

Usage is not a strategy. Token consumption is not a KPI. And a model release cadence that outpaces your evaluation cycle is not an advantage. It is a distraction.

The organizations that get this right in the next quarter will not be the ones that adopted the most. They will be the ones who measured the right things.

Context, as usual, is required.

James

Bench

• Microsoft Build June 2 to 3. Nadella’s Copilot positioning deserves close reading given the Claude Code cancellation backdrop.

• Webflow announced its second 8 percent workforce reduction under CEO Linda Tong, framed explicitly around agentic AI. The ClickUp pattern from last week is hardening into a category.

• Apple WWDC opens June 8 with Siri 2.0 and Gemini integration across 2 billion devices. The largest consumer AI deployment in history gets its public demo in seven days.

• SpaceX IPO roadshow begins June 4. The S-1 revealed Anthropic is paying SpaceX $1.25 billion per month for GPU compute. Infrastructure is still the actual market.

Anthropic is closing a $30 billion funding round at a pre-money valuation above $900 billion, the second $30 billion round it has closed in 2026. If it lands at the upper end, the Claude maker passes OpenAI’s $852 billion March valuation for the first time.

The number that should change how you plan is the one buried in the projections. Anthropic is forecasting $10.9 billion in Q2 revenue, up 130 percent from Q1, and its first quarterly operating profit ever. The story that frontier AI is structurally unprofitable is being quietly retired.

Plan for vendor pricing power to harden in the back half of the year. If your 2026 AI stack assumed Anthropic and OpenAI would keep competing each other into the ground on price, revisit that assumption now.

2. Google is turning search into an action layer

Google’s post-I/O messaging is now unambiguous. Search is no longer a place where people retrieve information. It is becoming a place where people delegate tasks. Nilay Patel’s Decoder interview with Sundar Pichai is the best single read on the shift. AI Mode now serves more than one billion monthly users. Organic click-through from queries featuring AI Overviews is down 61 percent since mid-2024.

For years, the operating assumption was to create useful content, earn attention, and capture demand. AI search compresses the middle of the funnel. The user may not click. The source may not be visible. The brand may only matter if its expertise is legible enough for an intermediary system to retrieve, summarize, trust, and act on.

If your 2026 plan still treats organic traffic as a top-of-funnel KPI, it was obsolete before the year started.

3. Practical: Notion just became an orchestration layer

On May 13, Notion launched its Developer Platform. Three pieces matter. Workers, a hosted runtime for custom code. An External Agents API that lets Claude, Codex, and Decagon operate as visible participants in your workspace. A Markdown API built for how agents actually think. Workers stays free through August 11 before moving to Notion’s credit pricing at $10 per 1,000 credits.

The competitive framing is direct. Notion is positioning itself against Zapier and Make as an agent orchestration layer, not as a document tool.

Before you ask whether your team needs AI agents, ask whether your team has agent-ready context. Clean ownership with up-to-date documentation that powers explicit workflows. Without those, the agent layer just accelerates ambiguity. The teams that use the next ten weeks to clean their workspace will be a quarter ahead of everyone else by October.

4. Hype Check: The AI ROI conversation got less polite

Uber’s president told The Verge that AI spending is getting “harder to justify” after the company reportedly burned through its annual AI budget four months into 2026. The specific culprit was rising token consumption on tools like Claude Code without a clear connection to improved engineering output.

That is the sentence every executive team should sit with. This is what happens when AI adoption is measured by usage instead of operating leverage. More tokens consumed does not mean more work is completed. More employees using AI is not a better process. More automation is not a healthier organization.

The trap is mechanical. A team gains access to powerful tools, usage spikes, everyone feels faster, the invoice arrives, the budget owner asks what changed, and the organization realizes it has been tracking activity rather than outcomes. AI programs do not fail only because the tools are bad. They fail because nobody designed the measurement system around the work.

5. Watch This Week: The agent story becomes an org-design story

TechCrunch reported on ClickUp’s mass layoff and framed it around the company replacing hundreds of employees with thousands of AI agents. Set aside the specific decision. The language is the story. Once leaders describe AI as a “teammate,” “agent,” or “digital worker,” they start making org-design decisions around it.

This is where most companies are underprepared. They are buying agentic systems before they have agentic management practices. Boring work often carries tacit knowledge. The person doing the repetitive task knows which client always needs a second look, which report field is technically optional but politically important, which “simple” workflow breaks every third Thursday because two systems disagree. Remove the person without capturing the context, and you do not get automation. You get operational amnesia with a better UI.

Watch the next three months for two patterns. Companies announcing layoffs framed around AI agents. And companies quietly rehiring six months later when the quality decay shows up in the customer data.

The companies that win with AI will not be the ones that adopt the newest tools fastest. They will be the ones who redesign the surrounding operating model. The workflows, measurement systems, and human review points that make the technology useful rather than theatrical.

AI adoption is no longer a tool-choice problem. It is a context problem.

And context, as usual, is required.

James

Bench (candidates for Notes this week)

• Trump pulled the AI executive order signing on Thursday, citing US-China competitiveness. The order would have established a voluntary 14 to 90 day pre-launch review for frontier models.

• OpenAI’s Grupo Folha and Grupo UOL partnership signals the labs are competing on content distribution and citation trust, not just model quality.

• Anthropic continues pushing Claude into regulated workflows: KPMG, PwC expanded partnership, Claude for Small Business, financial services agents.

• SpaceX’s IPO prospectus revealed Anthropic is paying SpaceX $1.25 billion per month through May 2029 for GPU compute. The infrastructure layer is the actual market.

Subscribe now