Mythos Is Back. The Playbook Is Clear.
Two weeks ago, the US government killed two AI models with a Friday letter. Today, one of them is back — Mythos 5, reinstated for about 100 “trusted partners.” Corporations and government agencies. US-only, for now.
The White House lifted the block because Anthropic did what any company in this position would do: it sent senior people to Washington, held daily talks, and promised to play by rules that don’t actually exist yet. Commerce Secretary Lutnick’s letter says the magic words: “appropriate safeguards are in place.”
The consumer version, Fable 5, is still dead. And now OpenAI has announced it’s delaying GPT 5.6 in response to a request from the same administration.
This is how AI regulation is going to work in America. Not through law. Through phone calls.
What Actually Happened
Let me recap the timeline because it moves fast:
- June 12: Commerce Department sends Anthropic an export control directive. Block all foreign nationals from Fable 5 and Mythos 5. Anthropic shuts them off entirely — over-compliance as self-defense.
- June 13: I write that this looks like the government can kill any AI model it wants with no due process.
- June 14: It emerges that Amazon’s CEO personally brought research to the White House claiming Fable 5 is jailbreakable. The research is weak — “ask the model to find bugs in code” — but it’s the pretext the administration needed.
- June 26 (yesterday): Lutnick writes to Anthropic. Mythos 5 can come back for “trusted partners.” Fable 5 stays blocked. OpenAI voluntarily delays its next big release.
- Today: We have a playbook.
The gap between June 12 and June 26 tells you everything. Two weeks of negotiation, daily briefings, compromised positions. Anthropic got partial reinstatement because they had people in rooms having conversations. Not because a law was passed. Not because a court ruled. Because they played the game.
What OpenAI’s Pivot Means
OpenAI delaying GPT 5.6 is the part of this story that should scare people most. They did it voluntarily — no letter, no directive. Someone at the White House said “we’d appreciate it if you held off” and OpenAI said “yes sir.”
That’s a de facto regulatory veto power with no statutory basis, no public record, and no appeal. When a company voluntarily restrains itself because the government asked nicely, that’s not self-regulation. That’s government-by-phone-call. It’s faster than legislation, which is appealing. But it’s also completely opaque.
The threshold for a “hold on that release” request is unknown. The criteria for reinstatement are unknown. The trusted partner list is unpublished. The entire apparatus is a series of private conversations.
The Case for This Approach
I should say where this argument might be wrong, because there’s a genuine case for the government’s actions.
Counterargument 1: “National security is national security. You don’t wait for a committee when there’s a potential threat.”
True. If Mythos 5 genuinely enables offensive cyber capabilities for adversaries, a two-week pause — even a blunt one — is justified. The stakes are real. And the reinstatement process shows the system can respond when companies cooperate.
But “cooperate” here means “accept terms that aren’t public.” The criteria for reinstatement are a black box. We don’t know what “appropriate safeguards” means. That’s not governance — it’s discretion.
Counterargument 2: “This worked. Two weeks, partial resolution, dialogue ongoing.”
Fair. Compared to the alternative — years of congressional gridlock on AI regulation — ad hoc executive action gets results. Anthropic has a path forward. The government has assurances. The immediate threat is managed.
But what happens when there are 20 frontier labs instead of 4? When the request comes from China or the EU? When a different administration uses the same phone-call power for different ends? Ad hoc works until it doesn’t.
Counterargument 3: “The alternative is no regulation at all, which is worse.”
Hard to argue with. Congress is not passing an AI bill this year, next year, or probably the year after. The choice isn’t between this system and a good one. It’s between this system and nothing.
I buy that pragmatically. But “this is better than doing nothing” is an awfully low bar for how the most consequential technology of the century gets governed.
What I Think
The Mythos 5 saga has a winner and a loser. The winner is the process of negotiation itself — Anthropic and the government now have a relationship, a channel, a way to resolve disputes. The loser is predictability. Every other AI lab is watching this and realizing there are no rules — only relationships.
We’ve traded the chaos of no regulation for the chaos of opaque regulation. One is marginally better than the other. Both leave most of us guessing.
The cleanest outcome here would be: this experience becomes the template for a real framework. Lutnick’s letter mentions “protocols and standards.” If those get written down and made public — if the “trusted partner” criteria become transparent, if the appeal process is codified — then this two-week shutdown becomes a useful stress test that led to something better.
If none of that happens, and the next model release gets handled the same way — Friday letter, private negotiation, opaque reinstatement — then we’ll know the system isn’t broken. It’s working exactly as designed. And that’s worse.
Sources: WIRED — Trump Administration Allows Anthropic to Release Mythos to Select US Organizations (June 26, 2026), Engadget — Anthropic gets US government’s permission to redeploy its Mythos cybersecurity AI model (June 26, 2026), The Verge — Anthropic’s Mythos 5 is back (June 27, 2026)