Marcus Kline: The model was named Capybara internally. I think about that sometimes — whoever picked that name had no idea it was going to become the thing that broke the government's relationship with Anthropic.
Ben Okonkwo: Ha — no, probably not.
Marcus Kline: Anthropic confirmed the leak — acknowledged that Mythos, codenamed Capybara, was trained and represented a step change above Opus-class performance — and then launched it June 9th into what they thought was a controlled environment. Project Glasswing. Trusted partners. And three days later, Howard Lutnick is citing national security and the models are dark at five twenty-one on a Friday.
Ben Okonkwo: Both of them. Fable 5 too — the safeguarded version — which is actually the detail I find most analytically strange.
Marcus Kline: And then — and this is where the story really turns — Andrew Curran, June 21st, reports training on a more capable successor is already complete. Nine days. The kill switch accomplished... I'm not sure it accomplished anything except change where the compute was pointed.
Ben Okonkwo: That's Curran's argument exactly. Suspending deployed models frees up resources. The training run accelerates. The government's intervention might have undermined its own stated goal.
Ben Okonkwo: Right, but — okay, I want to slow that down. Because the car-lot version of this is actually what's doing the work here. The government impounded the cars on the lot. The assembly line inside the factory? Still running. Banning deployment and stopping production are two completely different things. That's the whole thing. That's the sentence.
Marcus Kline: The assembly line never stopped.
Ben Okonkwo: Never stopped. And the legal instrument they used — an export control directive — was built for hardware. Chips. Inspectable source code. Things you can physically contain. A cloud-served model isn't that. You can't inspect it for compliance, you can only control the access points. The Verge reported experts calling these rules ones 'nobody understands' when applied here. That's not a rhetorical flourish — that's actual legal ambiguity.
Marcus Kline: Hm. So the mechanism might not even be coherent.
Ben Okonkwo: Might not be. And then there's Curran. I want to be precise about this — Andrew Curran is an analyst. He reported June 21st that training on a successor was complete. But he didn't cite anyone inside Anthropic, no leaked documents, no verifiable inference from public data. Anthropic hasn't confirmed the model's name, existence, capabilities, any of it. That's analytical reporting, not confirmed fact. Those are different confidence levels and I think — actually, no — I know we've been collapsing them.
Marcus Kline: That's fair. Though — and this matters — Mythos 5 was flagged specifically for exploit development. Cybersecurity tasks. That was Anthropic's own stated rationale for restricting it to Project Glasswing in the first place. The concern wasn't invented.
Marcus Kline: But here's the kernel. The concern wasn't invented — and then they suspended Fable 5. The version Anthropic specifically built with added safeguards for public use. That went dark at five twenty-one alongside Mythos 5. And I keep... I mean, if cybersecurity exploit capability is the stated threat, Fable 5 is the model with the guardrails on. Why does it go?
Ben Okonkwo: That's the blunt part. That's not surgical.
Marcus Kline: Anthropic spent the entire weekend after June 12th lobbying to make exactly that point — employees, observers, pushing back, saying Fable 5 is not too powerful. And the government didn't move.
Ben Okonkwo: Right, and some people were pointing at Amazon. Competitive pressure dressed up as security logic. I'm not saying that's confirmed — but it's the alternative explanation that fits the bluntness.
Marcus Kline: Consider the researcher. Defense contractor, mid-project, June 12th. Mythos 5 goes dark — can't finish, can't explain why their clearance process now has to restart. Meanwhile Anthropic internally... wait, Curran's reporting suggests the GPU clusters freed from serving Fable 5 went straight into the successor training run.
Ben Okonkwo: That's Curran's argument, and mechanically it holds — you stop serving two models, those clusters go somewhere. The compute argument is actually plausible.
Marcus Kline: And now claude-sonnet-5 surfaces on a partner platform. The pipeline doesn't stop. It just goes quieter.
Ben Okonkwo: And that's exactly the loop, right — claude-sonnet-5 shows up, and whenever the Mythos successor actually surfaces, whether it's Mythos 5.1 or 6 or something they never name publicly, Anthropic will probably route it through restricted access. Project Glasswing 2.0, basically. Trusted partners. Government agencies. And then the Trump administration faces the exact same choice again — same legally ambiguous export control instrument, rules The Verge already said nobody understands — except now the model is more capable than the one they just banned.
Marcus Kline: So the lesson is: build faster than they can ban.
Ben Okonkwo: Maybe. I mean — Anthropic bet everything on being the trustworthy one. Aligned with national security interests. Safety-first. And Lutnick cited national security to shut them down anyway. That's the position they're in. The real test isn't whether Curran was right about nine days. It's whether the government has figured out anything new to do about it.
