Anthropic’s Claude Fable 5 — the most capable model they’d ever shipped — was live for about three days before someone switched it off. Our boy got the “always has been” treatment. Here’s the full breakdown; the short version is below.
Two things matter. What actually happened (a lot of it is still fog of war), and the part that matters more: how to set yourself up so the next time a frontier model blinks out, you’re not dead in the water.
What actually happened
Anthropic launched Fable 5 on Tuesday, June 9th — their most capable widely released model, built, in their words, for long-horizon agentic work: big multi-step coding and reasoning jobs. We got about three days with it.
Then Friday hit. Anthropic says that at 5:21 p.m. Eastern on June 12th it received a directive from the US government citing national security — per the Wall Street Journal, routed through Commerce Secretary Howard Lutnick. And it wasn’t narrow. Anthropic says it was ordered to suspend access for “any foreign national whether inside or outside the United States,” including its own foreign-national employees — which, by its account, locked some of its researchers out of their own models. To comply, Anthropic says it had to “abruptly disable” Fable 5 and Mythos 5 for every customer. Some people got hit mid-project. Light switch off. Was it a little dramatic? Sure — there’s a whiff of malicious compliance to the whole thing. But the models went dark, and that’s the part that matters.
Fable will be back — no date yet. Polymarket has it around 80% before July 1st, though the same market gives a 10% chance the US confirms aliens by year’s end, so price that confidence accordingly. By the time you read this, it may already be back up.
Why it happened — here’s where the fog descends
Anthropic says the government never handed it specific written details, but its understanding is that this came down to a “narrow, non-universal jailbreak”: asking the model to read a codebase and fix software flaws. Anthropic’s position is that the vulnerabilities looked relatively simple, that other public models — including OpenAI’s GPT-5.5 — could surface them just as easily, and that pulling a model used by millions over that is a massive overcorrection.
Per the Wall Street Journal, the spark was Amazon. Its researchers reportedly used a specific sequence of prompts to get Fable to surface vulnerability information across at least four different programs — the sort of thing it’s supposed to refuse. Amazon CEO Andy Jassy took it to senior officials, the White House convened, and officials asked Anthropic to fix or pull the model.
Security folks who’ve read the report are skeptical it rises to “dangerous.” GreyNoise founder Andrew Morris told the Journal it’s “a long way” from genuinely dangerous, and that the whole thing boils down to: Fable could find bugs, therefore — by the magic of transitive property — it could plant them too. Except, per the Journal, there’s no evidence Amazon’s researchers actually got there. Katie Moussouris of Luta Security, who read the report, put it flatly: “I’ve seen the paper. It’s not a jailbreak.”
There’s a second theory — export restrictions, on fears that Mythos or Fable could be distilled by a China-linked group. I’m skeptical of that one: Fable is a public model available internationally, and Anthropic already caught flak for putting unusually strong anti-distilling guardrails on it. Possible? Maybe. A little on the clickbait side? Also maybe.
On the politics, a lot of it traces to David Sacks — former AI and crypto czar, now a co-chair of the president’s science council, and a vocal Anthropic critic. Sacks said on X that a partner trusted by both sides found a jailbreak testing Fable, that the administration asked CEO Dario Amodei to fix or pull it, and that Amodei refused. Anthropic pushes back hard: it says it got 90 minutes’ notice and no real detail, and Amodei called the move “retaliatory and punitive.”
The wrinkles pile up. Amazon — reported to have raised the alarm — is also one of Anthropic’s biggest backers, recently adding another $5 billion on top of $8 billion. OpenAI, with GPT-5.6 looming, arguably benefits most from a hobbled Fable (though I suspect 5.6 may sit in the wings a little longer; launching into this opens a risky door). And there’s existing bad blood: Anthropic got tagged a “supply chain risk” earlier this year.
To be fair to the other side: if Mythos — big daddy Fable — really could hand anyone in the world a working cyber weapon, a government stepping in fast isn’t crazy.
The real takeaway
This isn’t “Anthropic good” or “Anthropic bad,” or Dario right or wrong. It’s that a tool this capable got flipped off like a light switch, no warning, while people were mid-project. As these models get woven into our work — and our creative, medical, and personal lives — that’s a big deal. The model you built your entire workflow around can vanish in a snap.
So when Fable comes back, or when OpenAI drops 5.6, don’t just spend the next frontier model on another one-off project. Use it to build something that makes you harder to knock over.
Build a second brain you actually own
Was Fable good? Honestly, yes — but not for the “I one-shotted GTA 6” posts. It was the boring stuff: planning and execution. I’ve been back on Opus 4.8 and GPT-5.5 in Codex for a few days, and both are genuinely good models that still tend to wander off on bigger jobs — Opus in particular loves a side quest. Fable finally walked me through a website build that had been on my list forever — a spot where every video I’ve made about, say, Kling lives in one place (about 400 of them to chew through). Plenty left to do, but it’s up. It also, unprompted, turned the OpenAI Sora page into an in-memoriam RIP page, which I found genuinely funny.
Here’s the one I’d build first, though. When Fable’s back — or when 5.6 lands — build your own second-brain OS. Not in a chat window. Use Cowork or Claude Code to do it (Codex if you’re on OpenAI’s side). Same spirit as the AI film pipeline from the last video: it’s just a folder on my desktop. Mine’s called AI Brains.
Structure it however you need. Mine has an overseer module — “Friday” — sitting on top of every project I’ve got running, feeding into Notion as a single source of truth so she doesn’t forget things and actually behaves like an assistant: a YouTube brain, the newsletter, personal filings, my AI film templates. I’m terrible with Notion and spreadsheets; I don’t touch any of it. Friday does. There’s even a little desktop “cockpit” widget for everything I need to do over the next month.
The trick is that loading Friday doesn’t load all of that — she just knows where it lives, so you’re not burning half a million tokens to say hello.
And the real point: because it lives in a folder as a stack of markdown files, I can point any model at it — frontier or not — and be running in no time. I can even use dumber models, because Fable built the thing elegantly (off about eight earlier versions I’d iterated with other models). That’s what I’d build first: take an older version or make a new one, and build the architecture for whatever you need to get done. The result doesn’t need Fable to run. You’re building a bridge you can drive any car over.
Keep a backup model on the shelf
This whole mess is a fair call to arms on the open-source front — with one caveat. You’ve probably seen claims that the right mix of open models gets you a “Fable-level model” running locally. That’s not true. Most open-source models right now land around Sonnet level in the Claude ecosystem — which is not nothing. And the fact that something like Google’s Gemma 4 runs on machines that aren’t a 5090 is a real deal. Install one. The day corporate or governmental bickering leaves you dead in the water, you’ll be glad it’s there.
If you want a full walkthrough of wiring all this together, say the word — happy to make that video.
The bottom line
Things are going to get weirder and bumpier over the next year. I’m not going full AI prepper here, but having your bugout folder ready to go is not a bad idea. I’ll let you know the second Fable is back online.
This video is sponsored by invideo’s Agent One. I came in with a loose idea and walked out with a 40-second cinematic sketch — Census, a little UFO noodle — and the whole time I felt like I was directing, not fighting the tool. It builds a visual look with you, takes direction while another generation is still cooking, and quietly maintains a context folder of your project — characters, color palette, references — so nothing gets lost. If you want to make a film, not just generate clips, give it a spin. Try invideo Agent One here.
Watch the full breakdown: