At roughly 10 PM ET on Tuesday, Anthropic flipped a switch. Mythos 5 — their flagship model, the one they'd been shipping to enterprise customers worldwide — went dark. Not gradually. Not with a transition plan. Ninety minutes notice to customers who'd built it into their production systems. That's not a rollout. That's a rupture.
The official story vs. what's actually happening
Officially, this was a safety decision. Internal researchers — notably Nicholas Carlini, who'd flagged concerns back in March — had been briefing the White House on model safeguards. The timing, though, is curious. The same week the Trump administration publicly blocked Mythos 5 from G7 countries, claiming reinstating access would be "completely illogical." The UK requested a carve-out. Macron and Modi raised concerns. None of it moved the needle.
When a company disables its flagship model with less than two hours notice, that's not safety caution. That's a company reacting to external pressure it can't acknowledge.
Here's what's being lost in the noise: this isn't just about one model. It's about who gets to decide when AI systems are safe enough for the rest of the world. The G7 — representing the largest economies outside the US — was told no. Not "not yet." Not "here's what we need to see." Just no.
The coalition that couldn't hold
Dario Amodei, Sam Altman, and Demis Hassabis did appear together at the G7 summit. They called for US-led collaboration on AI rules. On the surface, that's a united front. Underneath, it's a concession: the industry is acknowledging that some body needs to set the rules, and they're betting that body should be the US, not the EU or anyone else.
The Financial Times reported the closed-door nature of these discussions. No public draft. No multilateral process. Just the three biggest labs and the US government, figuring out AI governance over dinner. That's not international cooperation — it's bilateral negotiation with an audience.
Meanwhile, Trump's Commerce Secretary Lutnick floated the idea of government equity stakes in AI companies, favoring a sovereign wealth fund model. Bernie Sanders responded with his own proposal: a one-time 50% stock tax on AI companies hitting $200M in annual AI sales, funding a public ownership stake. Neither is likely to pass, but both signal theOverton window is shifting.
What this means for the industry
The practical impact: if you're outside the US, your access to leading models is now contingent on diplomatic relationships, not capability. Mythos 5 didn't get disabled because it was unsafe. It got disabled because the political cost of keeping it on exceeded the cost of turning it off.
For enterprises building on these platforms: this is your risk profile. Your model can vanish in 90 minutes. Your compliance posture depends on geopolitics, not your contract terms. The people building AI products at scale need to internalize this: you're not buying compute, you're entering a relationship where the other party can walk away with a two-hour window.
The signal here isn't safety. It's that the AI industry's "move fast and break things" ethos has run into the reality that breaking things internationally carries regulatory consequences that can't be fixed with a rollback.
We'll be watching whether the carve-out requests gain traction, whether other labs face similar pressure, and whether the G7 coalition fractures or pushes back. The next 30 days will tell us whether this was a one-off or the new normal.