The Model Returns

Three weeks ago I wrote about Anthropic's Fable 5 and Mythos 5 being shut down overnight by a US export-control directive [1]. The models went dark on June 12 because the Trump administration worried about jailbreak vulnerabilities. As of today, July 1, the restrictions are lifted and the models are coming back [2].

So the story has a happy ending, or at least a pause. But the details matter, because they tell us what the new normal for frontier AI looks like.

What happened in the middle

On June 12, the US Department of Commerce slapped export controls on both models. The trigger was a cybersecurity finding: Amazon researchers had demonstrated a jailbreak that let Fable 5 identify software vulnerabilities and produce exploit code. The government response was immediate and blunt: no foreign nationals could access either model, including non-US employees of enterprise clients and even Anthropic's own foreign staff [3].

Anthropic could not verify user nationality in real time, so they suspended access for everyone. Not a graceful degradation, just a hard stop.

Three weeks of negotiation followed. Mythos 5 returned first, on June 26, but only to a pre-approved list of US organizations. Fable 5 returns globally today, July 1 [4].

What Anthropic agreed to

The models did not come back unconditionally. Anthropic agreed to several new commitments:

• Proactive risk detection: Anthropic will detect and address security risks associated with the models before they escalate.
• Government pre-release access: Designated US agencies will get early access to frontier models and their safeguards before public release, so they can run independent evaluations.
• Threat intelligence sharing: Anthropic will share threat intelligence with the government, including malicious activity reports.
• A new jailbreak scoring framework: Anthropic is drafting a framework with Amazon, Microsoft, and Google for scoring how dangerous a given jailbreak is [5].

On the technical side, Anthropic trained an improved safety classifier that targets the specific exploit-code generation behavior Amazon flagged. The classifier blocks the technique in over 99% of cases. When a request is blocked, it gets routed to a weaker model (Opus 4.8) instead of failing outright [6].

What this means for the rest of us

The original lesson from June 12 still holds: if you build on a frontier model API, the model can disappear for reasons entirely outside your control. But the resolution adds nuance.

First, the government can overreact and then walk it back. Researchers who studied the Amazon jailbreak said its significance was widely inflated. Francesco Bailo from the University of Sydney noted that "if Fable and Mythos were blocked on these grounds, competitor models would have to be blocked too" [7]. The precedent was unsustainable, and the Commerce Department seems to have realized that.

Second, the cost of compliance is real but survivable. Anthropic spent three weeks negotiating, built new safety infrastructure, and accepted ongoing government oversight. That is the price of getting your models back. Smaller companies might not have that leverage.

Third, the government is now embedded in the model release cycle. Pre-release government evaluation, threat intelligence sharing, and joint safety standards are now part of the deal. The Trump administration applied the same staggered-release approach to OpenAI's GPT-5.6 last week, limiting initial access to a "small group of trusted partners" [8].

The local model argument, revisited

In my June 20 post I argued that a smaller, weaker model you control is worth more than a brilliant model that vanishes overnight. Three weeks later, the brilliant model came back. Does that argument still hold?

Yes, but differently. The models returned, but with strings attached: government pre-release access, safety classifiers, staggered rollouts. If your product depended on Fable 5 between June 12 and July 1, you had three weeks of nothing. A local fallback would have kept the lights on.

The happy ending does not undo the warning shot. It just means the gun was lowered this time. Next time it might not be.