Fable 5 is back with the government in the loop

The fable got a sequel, and a lot faster than I'd have bet. Three weeks after pulling Fable 5 on Washington's orders, Anthropic has switched it back on. On the evening of June 30, the Commerce Department lifted the export controls on Fable 5 and Mythos 5. As of July 1, the consumer model is available worldwide again¹. Am I glad to have it back? Honestly, yes. But I ended my last article on a question, how long a country can afford not to have its own model. The news answers with a trickier one first. On what terms did it come back?

The story resumes, with strings attached

The power isn't back for everyone. Fable 5, the model anyone can use, is back worldwide. Mythos 5, the guardrail-free version for security professionals, only reopens to a circle of approved US organizations, with the government's green light since June 26¹. So the muzzled model goes global again, and the unmuzzled one stays in American hands. The sovereignty I wrote about last month didn't take long to make its point.

To keep it from happening again, Anthropic also relaunched Fable 5 with a new set of classifiers² meant to spot and block cybersecurity requests it considers sensitive. Except the net is still pretty coarse. Tasks as ordinary as writing or debugging code get bounced, for now, to Opus 4.8, the previous generation's model¹. The company itself admits its classifier catches plenty of perfectly legitimate requests along the way. It promises to fine-tune it over the coming weeks to tell real misuse from a developer's plain curiosity.

The accuser was Amazon

Last month, I summed up the affair with what little we knew, a "rival company" claiming to have gotten around the model's protections. The veil is off, and the name is worth a grin. It was Amazon. Its researchers showed that handing Fable 5 a codebase could get it to spot a string of security flaws and even, in one case, write a rough start on code to exploit them¹. Amazon's CEO himself, Andy Jassy, reportedly alerted the White House³.

Washington's version is less flattering to Anthropic. Adviser David Sacks accuses the company of refusing to fix the flaw before the controls came down, while a Chinese group had supposedly already gotten its hands on the model, against a backdrop of ongoing tension with Beijing⁴.

And Amazon isn't just any competitor. It's Anthropic's biggest backer, having poured billions into the company while building its own models on the side. Watching it trigger, with one note to the authorities, the shutdown of the crown jewel it partly owns is quite a balancing act. The punchline is even better, since that same Amazon is now among the partners writing the framework for judging how serious a jailbreak is. The arsonist joined the fire department.

The best part is that it proves my June hunch right. Anthropic now says Amazon's demo was nothing special. Every model tested gave the same result, from little Haiku 4.5 through the successive Opus releases, by way of OpenAI's GPT-5.4 and 5.5 and China's Kimi K2.7¹. So what Fable got blamed for applies to any model that can read code, because reading code also means spotting its flaws. They unplugged a tool for a skill every serious model on the market has.

The state moves upstream

The comeback came at the price of a lasting closeness between the company and the government. Anthropic is expanding its work with the US government on evaluating its models. In practice, the government gets access to models and their safeguards before release, gets quick reports on jailbreaks and misuse, and joins in shared research¹.

That goes way beyond last month's incident. The state no longer just steps in after the fact, unplugging what worries it. It's moving upstream, into the engine room, where it gets decided what a model is allowed to know how to do. The "state secret" I wrote about becomes a quiet, permanent partnership between a private company and its government.

And it doesn't stop at Anthropic. With Amazon, Microsoft, and Google, the company is sketching a shared framework for rating how serious a jailbreak is, based on four criteria running from the real boost it gives an attacker to how easily it can be used¹. It's also opening a program on HackerOne, a platform that pays security researchers to report flaws. The intent is good, and it'd be petty to scoff at an industry finally trying to set some rules. But those rules are being written by the same handful of American companies, under the eye of the same government, for models the rest of the world will just get to use.

Sonnet 5, the lesson learned

The day before, on June 30, Anthropic released Claude Sonnet 5, a lighter model aiming for Opus 4.8's performance at a much lower price⁵. Given the timing, what stands out is that it deliberately underperforms on cybersecurity, with its safeguards on by default. June's lesson stuck, and it became a product feature. A model built not to know how to do the thing that might get it unplugged⁶.

So Fable 5 is back, and I'd be lying if I said I wasn't happy about it. But the fable changed its moral along the way. It no longer just says a government can switch off a model with the stroke of a pen. It says a government can now pull up a chair at the table where the model gets built, and stay. The power's back on. The question is who's holding the switch.

Notes

1. Anthropic, "Redeploying Claude Fable 5", June 30, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

2. A classifier is a small model whose job is sorting, here filing each request as "harmless" or "suspicious" before it reaches the main model. A bouncer, basically, who also turns away perfectly decent people for wearing sneakers. ↩

3. Fortune, "How a warning from Amazon led the White House to shut down Anthropic's Mythos model", June 14, 2026. ↩

4. Tom's Hardware, "Trump adviser David Sacks says Anthropic refused to fix Fable 5 jailbreak before US export controls", June 2026. ↩

5. Anthropic, "Claude Sonnet 5", June 30, 2026. ↩

6. Like the kid who flunks his chemistry test on purpose so he won't get sent to the principal's office. We all knew one. ↩