The Model I Praised Yesterday Is Gone Today

YunLab · Industry Watch

Yesterday I published a piece recapping 35 hours of work with Fable 5 — 52 commits, four usable things moved a big step forward. Less than a day after it went up, Fable 5 is gone.

I'm writing this one on Opus 4.8. Because the model I praised yesterday has been pulled by the US government.

Three days

It happened fast. On June 9, Anthropic publicly launched Fable 5 — the first time it opened its most powerful "Mythos-class" model to ordinary users. On the afternoon of June 12, a US government export-control directive ordered the suspension of all access to Fable 5 and its sibling model Mythos 5 for any foreign national — including the company's own foreign-national employees. Anthropic disabled both models for every customer worldwide that same day. Not just foreign users — everyone — because it couldn't reliably pick out foreign nationals one by one on a shared cloud service in real time.

From launch to shutdown: three days.

(One thing to get straight: reporting points the directive at the US Commerce Department, but Anthropic's official statement says only "the US government, citing national security authorities" — it names no agency. I'll go with the official wording: "the US government," not some confirmed department.)

The government says someone jailbroke it

The reason given: the government believes someone found a way to "jailbreak" Fable 5. Anthropic says it reviewed the demonstration — the so-called jailbreak amounts to asking the model to read a codebase and find the software flaws in it, which surfaced "a small number of previously known, minor vulnerabilities." Anthropic stresses that this capability exists in other models too (including OpenAI's GPT-5.5) and is something cybersecurity people use every day.

Here I have to flag something honestly: "small, previously known" is Anthropic's own characterization, made in the context of disputing the directive. The government sees the same thing as a national-security risk. The two sides' judgments of "how serious is this, really" point in opposite directions.

Anthropic's stance is clear: comply, but object. In its own words — "We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people." And a heavier line: if this standard were applied across the industry, it "would essentially halt all new model deployments for all frontier model providers."

One distinction matters: this is a suspension, not a discontinuation. Anthropic says it believes this is a misunderstanding and is working to restore access as soon as possible — it just hasn't given a timeline. Apart from these two models, every other Claude model — including the Opus 4.8 I'm typing this on — is completely unaffected.

What this has to do with someone who just uses models to get work done

I don't build models. I'm a middle-aged guy who uses them to get work done. But this taught me a solid lesson.

In yesterday's piece I wrote a line: its memory is external, it doesn't make the calls for me, and leverage amplifies output as much as it amplifies risk. Today I have to add one more — this lever isn't mine to control, and it isn't even fully the model vendor's to control. A tool I ran 52 commits on yesterday can vanish today over a single letter, in three days. It didn't crash, and I didn't misuse it. Someone flipped a switch in a place I can't reach.

So I'm more sure of two things now.

• Don't build your house on a single model. My system — memory files, task folders, handoff docs — is model-agnostic. Fable's gone today, I swap in Opus; Opus has trouble tomorrow, I swap in something else; my workflow connects either way. What's actually mine is the engineering system that lets any model plug in — not any one specific model. The model is rented. The workflow is owned.
• Local capability, the kind held in your own hands, is worth more and more. Those MLX local models on my machine, that offline foundation — I used to treat them as a side hobby. Now they look like insurance. The most powerful thing in the cloud might just not be there one morning; the slower, dumber thing running locally at least can't be switched off by someone across the Pacific.