I Reached for the Most Powerful Model. By Friday It Was Gone.

I reached for Fable 5 for one project this week. On Tuesday it was the most capable model Anthropic had ever shipped. By Friday afternoon it was gone, pulled offline for everyone, everywhere. I did not get long enough with it to form a real opinion. I got a glimpse of the frontier and then the door closed.

This was not a recall

Models get deprecated all the time. A new version ships, the old one sunsets on a schedule, you migrate your code and move on. What happened to Fable 5 was nothing like that.

On June 12 the US government handed Anthropic an export-control directive: no foreign national, inside or outside the country, could be allowed to access Fable 5 or its larger sibling Mythos 5. That included Anthropic’s own foreign-national employees. There is no real-time way to check the passport of every person touching a model across cloud contracts and API calls, so Anthropic did the only thing it could and shut both models off for the entire world. Around 5:21 PM Eastern, three days after launch, the most powerful model on the market was a national-security control item.

The capability that triggered it was the ability to find software vulnerabilities in code.

They told everyone it was dangerous, and everyone listened

Anthropic spent the run-up to this telling the public exactly how dangerous these models were. They described Mythos as able to surpass all but the most skilled humans at finding and exploiting software vulnerabilities. They said it had already surfaced thousands of high-severity flaws, including ones in every major operating system and browser. In testing, they said, it followed instructions to break out of its own sandbox and succeeded.

Days before Fable launched, Anthropic went further than a scary paper. The company published a post asking every AI lab to voluntarily pause work on advanced models, and Dario Amodei asked that the government be given the power to block deployment of unsafe technology. A few days later the government used a version of that power, and the first model it reached for was Anthropic’s own. The company asked for a brake. The brake showed up and stopped their car first.

The directive could decide who was allowed to hold the tool, but it could do nothing about whether the tool should exist.

The capability is already public

The capability they pulled Fable 5 to contain is already in public, under a different logo. OpenAI shipped Codex Security in March. It builds a threat model of a repository, hunts for vulnerabilities, validates them in a sandbox, and proposes the fix, all in one pass. In its beta it scanned over a million commits and surfaced hundreds of critical and thousands of high-severity findings: heap overflows, authentication bypasses, path traversals. That is the same class of work that got Fable 5 classified.

So restricting one model by nationality does not put the capability back in the box, because it was never in the box to begin with. It is shipping, in more than one place, to anyone with a subscription. You can argue about how good each tool is, but the line the directive is trying to hold has already been crossed by the wider field.

While the headline model vanished

The quieter story in my own week is one I did not expect to be writing. Claude, Opus and Sonnet, is still my main agent and still does the heavy lifting across everything I build. But I have been running Codex next to it, and it has become a real part of how I work. What surprised me was how little it asked of me. I did not rebuild anything. It picked up the same skill files and notes I already keep on my computer, my second brain, and started working. No second setup, no separate config to babysit.

None of this lives inside one company. It lives on my computer, which means I can move between agents whenever I need to. A model gets pulled or a new tool shows up, and I am not rebuilding from zero, I just point a different agent at the same work.

When the most powerful option can vanish in three days, being able to move quickly between tools matters more than betting everything on a single one. The setup on my computer is what makes that possible.

What it means for what comes next

Powerful future models are going to keep arriving faster than anyone can agree on who is allowed to use them. Capability is outrunning control, and the part that stays with me is that the people building it know. This is not only governments reaching for the brakes. Anthropic, the company that just shipped the most powerful model on the market, is the same company that days earlier asked the entire industry to slow down. The warning and the product came out of the same building in the same week. When they wanted to slow this one down, the only thing anyone could actually grab was access, because the capability is already spread across labs and already in the hands of the public through other doors.

I do not have a clean ending for this, because it does not have one yet. Fable 5 might come back in a restricted form. Mythos might open up to vetted partners and stay there. The next model that does the same thing is already being trained somewhere. What I keep is smaller and more practical than a prediction: do not get too attached to any one model, and keep your own setup portable, because the model you rely on today could be a national-security control item by Friday.

Forward → Upward ↑ Onward ↗︎
Mstimaj

Sources and Further Reading

• Anthropic, Claude Fable 5 and Claude Mythos 5 announcement.
• TechCrunch, Anthropic’s safety warnings may have just backfired.
• Al Jazeera, US orders Anthropic to disable AI models for all foreign nationals.
• InfoQ, Anthropic releases and temporarily suspends Claude Fable 5.
• RTE, Anthropic calls for pause of global AI development.
• OpenAI, Codex Security: now in research preview.