Anthropic suspended Claude Fable 5 and Mythos 5 for every customer worldwide on the evening of June 12, 2026, three days after Fable 5 went generally available. Not throttled, not degraded. Off. The cause was not an outage or a capacity wall but a US export-control directive citing national security, written so that no foreign national, inside the US or out, could be allowed near the two models. Unable to filter users by nationality in real time, Anthropic killed both models for its whole customer base. Here is the 72-hour timeline, the code-reading jailbreak that triggered it, Anthropic's pushback, and what to do if your API calls just went dark.
The short answer
Anthropic disabled Claude Fable 5 and Mythos 5 for every customer, worldwide, the
evening of June 12, 2026. A US export-control directive citing national security ordered
that no foreign national be allowed near the two models, and Anthropic says it cannot
pick foreign users out of the crowd in real time, so it killed both models for the whole
customer base. If your code targets claude-fable-5 or claude-mythos-5, fall back to
claude-opus-4-8. Still live, and it costs half as much.
Build on Claude? Your Friday night may have broken on you. Sometime after dinner on the US East Coast, calls to claude-fable-5 and claude-mythos-5 just started failing. Not your code. Anthropic had switched both models off. Not throttled, not degraded. Off. Everyone, everywhere.
And the why is the weird part. No outage. No capacity wall. A US government order, written in a way that left Anthropic almost nowhere to go except pull the plug for the whole planet. Below: what happened, how a rule about foreign nationals managed to kill your account in France or Brazil or Ohio, and what you actually do if your pipeline just went dark.
A 72-hour timeline
The speed is what gets me. These were brand new models. Fable 5 went generally available on June 9, across the Anthropic API, the claude.ai apps, Amazon Bedrock, Google Vertex AI and Microsoft Foundry. First public model in Anthropic's new Mythos-class tier, sold as frontier capability wrapped in extra safeguards so it could ship wide. We wrote up how it compared to Opus 4.8 on launch day, and honestly, it looked like a real step up.
Then June 12, 5:21pm ET. Anthropic says a government directive landed, citing national security authorities. By that evening both Fable 5 and the partner-only Mythos 5 were just gone. Three days. For a company that had spent the whole week bragging about how capable these things were, yanking them over a weekend? That stings.
Why a foreign-national rule shut off everyone
Here's the detail that trips people up. The order didn't say "turn off Fable 5." It said no foreign national may touch Fable 5 or Mythos 5, inside the United States or out. That's broad. Broad enough to sweep in tourists, visa holders, remote users abroad, even Anthropic's own staff who don't hold a US passport.
Now picture enforcing it. You'd have to verify the nationality of every single person behind every API key and every chat session, live, all the time, to a standard that satisfies an export-control regulator. No consumer AI platform is built for that. Pretending otherwise is the legally risky move. So Anthropic took the one clean option on the table: kill both models for everybody until this gets sorted. A rule aimed at some users turned into a kill switch for all of them, just because half-compliance wasn't a real option.
The trigger: a code-reading "jailbreak"
This is the part you'll care about most, because it sits right on top of normal security work. The government published no detailed technical case. What Anthropic describes getting was basically verbal evidence of a narrow, non-universal jailbreak: a way to get Fable 5 to read a specific codebase and identify software flaws in it.
Read that twice. If you do any application security, you just felt something. Pointing a capable model at a repo and asking "where are the bugs" isn't an exotic attack. It's a Tuesday. It's what static analysis tools and code-review assistants and half the AppSec industry already do, every single day, right out in the open.
The capability that got two frontier models recalled is the same one a lot of you reach for on purpose, with permission, to find and patch vulnerabilities before the attackers do. The line between "helpful code review" and "national security risk" just got a lot blurrier. Nobody drew it for us.
Anthropic says that when it actually looked at a demo of the technique, it turned up only a handful of previously known, minor vulnerabilities. Nothing exotic. The company also notes that other public models, OpenAI's GPT-5.5 among them, spit out similar stuff with no bypass needed at all. So the scary capability isn't unique to Fable 5, and the demo wasn't surfacing novel zero-days. It was rediscovering bugs that were already on the books.
Anthropic isn't taking it quietly
The company is complying. It also made a point of disagreeing, on the record, and the argument is worth hearing out. One narrow potential jailbreak shouldn't get a commercial product recalled, it says, any more than a single phishing email should get an entire email provider shut down. The same capability is everywhere anyway, used every day by legitimate cybersecurity pros. And then the line that should make every AI builder sit up: apply this standard consistently and you basically halt all new model deployments, because any model worth shipping can be talked into reading code and spotting flaws. That last point is hard to wave away.
Anthropic apologized to customers, said it's racing to restore access, and promised more technical detail within 24 hours of the shutdown. As of this writing? No firm restoration date.
What to do right now if you're affected
If your weekend just got interesting, here's the triage.
- Swap the model ID. Anywhere your code calls
claude-fable-5orclaude-mythos-5, point it atclaude-opus-4-8instead. Opus 4.8 is untouched, fully live, and runs at half the price ($5 / $25 per million tokens versus $10 / $50). The quality gap is small for most work. We said as much in our launch-day comparison. - Check every surface, not just the API. The suspension hit the Anthropic API, the claude.ai apps, Bedrock, Vertex AI and Foundry, all of it. Route through a cloud provider? Same model IDs, same dead end over there. A multi-provider setup won't save you this time.
- Mind your fallback logic. Built automatic retries, or a model-router? Make sure a failed Fable 5 call drops to Opus cleanly, instead of hammering a dead endpoint or coughing raw errors at your users.
- Don't hard-code the comeback yet. Access might be back fast. Might not. Make the model a config value, not a constant, so you can flip it back without a redeploy when the dust settles.
| Model | Status (June 13) | What to do |
|---|---|---|
claude-fable-5 | Suspended worldwide | Fall back to Opus 4.8 |
claude-mythos-5 | Suspended (was partner-only) | Fall back to Opus 4.8 |
claude-opus-4-8 | Live, unaffected | Use as the default for now |
| Sonnet / Haiku tiers | Live, unaffected | Fine for lighter workloads |
Why this matters past the weekend
Step back, and the outage is the least interesting bit. What got demonstrated this weekend: a government can, with one letter, pull a commercial AI model off the global market in hours. And the way export-control rules are built, a restriction meant for a few users can balloon into a shutdown for everybody. That's a precedent. Every company building on a frontier model now has to plan around it.
It also drags an old, awkward question back into the light. A capable model that reads code and finds bugs is dual-use by nature. Defender's best friend, attacker's force multiplier, and it's the exact same feature either way. Regulators are clearly starting to treat that as a controlled good. If you work in security, I think that framing reshapes which tools you're allowed to touch, and how, a lot sooner than most people expect. Could be wrong on the timing. The direction feels pretty clear though.
For now the boring advice is the right advice. Keep your model choice in config. Keep a working fallback, and don't assume any single model is a permanent fixture. This weekend proved one can vanish between one request and the next.
Sources
- Anthropic: statement on the US government directive to suspend access to Fable 5 and Mythos 5
- NBC News: Anthropic suspends new AI models after government directive
- CNBC: Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive
- Engadget: Anthropic blocks all customers' access to Fable 5 and Mythos 5
- The New Stack: Federal government orders Anthropic to pull Fable 5 and Mythos 5, three days after launch
Frequently asked questions
Are Claude Fable 5 and Mythos 5 gone for good?
No sign of that. Anthropic says it disabled them to comply with a government directive and is actively working to restore access, with more detail promised soon. No firm restoration date at the time of writing, so treat the suspension as ongoing, not permanent.
Why were the models cut off for everyone if the order only mentioned foreign nationals?
Because Anthropic says it cannot reliably separate foreign users from everyone else in real time. Verifying the nationality behind every API key and chat session, nonstop, to a regulator's standard? A consumer AI platform just cannot do that. So the company switched both models off for its whole customer base rather than risk non-compliance.
What was the security concern that triggered the order?
A reported trick that got Fable 5 to read a specific codebase and identify software flaws in it. Anthropic calls it a narrow, previously known issue that only surfaced minor, already-documented vulnerabilities, and points out that other public models produce similar output with no jailbreak at all.
My app calls Fable 5. What is the fastest fix?
Change the model ID to claude-opus-4-8. Opus 4.8 is untouched, costs half as much, and sits close in quality for most tasks. Keep the model name in a config value so you can flip back without a code change once Fable 5 returns.
Does this affect Opus 4.8, Sonnet or Haiku?
No. The directive named Fable 5 and Mythos 5 specifically, nothing else. The rest of the Claude lineup, Opus 4.8 included, stayed live and untouched across the API and the cloud providers.