What it means for us

Anthropic launches Fable 5 — then the U.S. government shuts it down

Anthropic launched Claude Fable 5 last Monday — its most capable model yet, priced at $10 per million input tokens and $50 per million output (a Tokensaurus). By Friday, the U.S. Commerce Department had issued an order requiring Anthropic to suspend all access for any foreign national, whether inside or outside the United States. Because Anthropic had no mechanism to apply that restriction selectively, it disabled both Fable 5 and its more restricted sibling Mythos 5 globally — for everyone. The company said it believed the action was based on a narrow, non-universal jailbreak technique and called it "a misunderstanding." It said it was working to restore access as soon as possible.

What does this mean for us? This is more than a question about the government's role vis-à-vis tech – think of it in terms of "what does your organization say when a tool you built workflows around disappears on a Friday afternoon" question. Granted, it was a bit early for Fable-based workflows, but it's never too early to think about your backup plan if your favourite model goes kaput.

Amazon reportedly flagged concerns — before the crackdown

Reuters reported that Amazon CEO Andy Jassy raised concerns with U.S. officials about security risks in Anthropic's most advanced models before the government acted. A public jailbreak demonstration shared on X on June 10 reportedly reached officials and helped trigger the directive. Anthropic is a company Amazon has invested heavily in. So: a major investor flagged a risk, regulators acted on it, and the company whose product got pulled found itself publicly disagreeing with the very oversight it had called for days earlier.

Microsoft restricted Fable before the government did — over data retention

Even before Friday's shutdown, Microsoft had already quietly blocked internal employee use of Fable 5. Not for capability reasons, but because of the data it retained. Anthropic's Mythos-class models retain prompts and outputs for up to 30 days for trust and safety monitoring — and Microsoft's concern was that confidential and customer data could be exposed.

What does this mean for us? Comms teams handle material that has no business sitting in a model's retention queue! Crisis drafts, layoff messaging, executive remarks for earnings calls, regulatory language. A better model is not automatically a better workflow if the data terms create legal or trust exposure. Before your team puts anything sensitive into any AI tool, ask one question: what's the retention policy?

G7 leaders invite AI executives to Évian

AI executives from Anthropic, OpenAI, Google DeepMind, Mistral AI, Cohere, Meta, Salesforce and Synthesia are expected at the G7 summit in Évian-les-Bains, beginning today (running June 15–17). The French presidency has made protection of minors online a priority alongside AI regulation, infrastructure and network sovereignty. The guest list reads less like a summit side event and more like a summons.

When heads of government sit down with AI executives at a G7 table, the communications frame is no longer innovation. It's accountability, sovereignty, child safety. The standard line of "We're committed to responsible AI" isn't going to cut it — or shouldn't. Every CEO with an AI strategy now needs to be able to say specifically what they're building, what they're declining to build, and why the public should trust the difference.

Anthropic calls for stronger federal AI rules — then objects when the government uses them

Earlier this week, Anthropic urged Congress not to pre-empt state AI laws without a strong federal framework in place. By Friday, the company was publicly disputing the way a federal directive had been applied to its own flagship model. Both positions may be entirely sincere. But audiences notice the gap.

If your organization says it supports responsible AI governance, it needs language ready for the day governance arrives at the door with a clipboard. "We support oversight" is not a communications position. It's an opening sentence. The paragraph that follows is what the speechwriter has to write.

The Tokenpocalypse: AI budgets are blowing up across the enterprise

Uber burned through its entire annual AI budget by April and capped engineers at $1,500 a month per tool. Walmart started rationing tokens after previously offering unlimited access. Microsoft cancelled most of its internal Claude Code licences. Meta tracked employees consuming 73.7 trillion tokens in a single month — on a leaderboard it's now dismantling. OpenAI's Sam Altman publicly acknowledged that companies are telling him they "know there's a ton of waste."

The "use AI everywhere" era is giving way to the "use the right AI for the right work" era — which is actually the correct answer, and not a retreat. The internal communications challenge is real: employees who were told to experiment freely are now being told to work within limits. That message needs to sound like operational maturity. The line that works: "We're not using less AI. We're using better-matched AI — the right model, for the right work, at the right cost."

And if any of this sounds familiar – you must have read my Substack article last week.

OpenAI and Anthropic keep warning about the future they're racing to build

Both companies continue issuing caution about the pace and risk of advanced AI development — while simultaneously releasing stronger models, expanding enterprise access, and heading toward public markets.

It's like the fire chief trying to sell you fireworks.

The more useful observation: risk language without corresponding behaviour is a liability, not a safety position. Would your organization's AI caution still sound credible if someone compared it line-by-line with your actual rollout plan? If the answer is "probably not," that's a comms problem to solve before an analyst or journalist gets to it first.

The real-time interpreter in your pocket (or: why the hell did I take all those French courses?)

Google launched Gemini 3.5 Live Translate on June 9 — a speech-to-speech translation model that works continuously in more than 70 languages, staying just seconds behind the speaker rather than waiting for a full sentence to finish. It preserves the speaker's tone, pitch and pacing. It's described as no setup, no turn-taking, no awkward silence while the app catches up. It's live in the Google Translate app on Android and iOS now, with Google Meet integration coming later this year for Workspace customers.

Okay, fine – that’s how it’s described. I have run this a few times, using a simple YouTube BBC Mundo news story and, well… impressive. However, it does take a few clicks to figure out how to get things going. Also, plenty of judder – which may be the limitations of my Mac more than the feature itself. Nevertheless, the benefits of this model are clear.

For communicators working internationally — which at this level is most of us — this could change the quality bar for on-the-ground interpretation. It shouldn’t replace a skilled human interpreter for a high-stakes keynote. But for a bilateral hallway conversation at a summit, a pre-meeting briefing in a second language, or a client debrief where someone's English is a third language they're working hard in, this will definitely be useful…and it’s “free” (you’ll need a Google account).

My only misgiving? Laws around recording conversations via phone or otherwise. I’m getting flashbacks to my media law classes.