Claude Fable 5 Is Dark: What To Do Right Now (June 2026)

Hot take: if your production stack broke on Friday because Anthropic’s Fable 5 and Mythos 5 went dark, that’s not Anthropic’s fault, and it’s not the government’s fault. It’s a fallback problem. The crackdown is a free stress test – and a lot of teams just failed it.

Here’s what actually happened, what’s still live, and the exact migration most teams should run today.

The 72-hour timeline

Anthropic shipped Fable 5 on June 9, 2026. Three days later, both Fable 5 and Mythos 5 were offline worldwide. Amazon CEO Andy Jassy was on a pre-scheduled June 11 call with White House officials about an unrelated topic when he raised concerns about Fable 5 vulnerabilities – that call set the chain of events in motion.

Commerce Secretary Howard Lutnick sent the directive at approximately 5:21 p.m. ET on June 12. It extended to foreign nationals both inside and outside the United States – including Anthropic’s own foreign-born employees. The company’s response: disable both models globally, because it had no way to enforce nationality-based access selectively.

That last detail is the whole tutorial in one sentence. Other Claude models still work.

What’s still live (as of June 2026)

If your code was calling claude-fable-5 or claude-mythos-5, it’s returning errors. Everything else in the Claude family is fine. Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5 are all online and unaffected – only Fable 5 and Mythos 5 were suspended by the June 12 directive.

# Before (broken)
model = "claude-fable-5"

# After (works today)
model = "claude-opus-4-8"

That’s it for most people. The migration is usually a single model-string change followed by re-running your evals. Don’t skip the evals step – Fable 5 and Opus 4.8 are different models, so any prompt tuned tightly to Fable’s behavior may need a temperature or system-prompt adjustment.

Why “just switch models” is the wrong instinct

Swapping the model string fixes your code. It does not fix your legal exposure.

The Greenberg Traurig legal advisory is blunt: switching to another model addresses continuity but doesn’t eliminate the need for a compliance review. Companies should map which users, business units, and jurisdictions had access to Fable 5 or Mythos 5 – including whether any foreign nationals inside or outside the U.S. could reach those models directly or indirectly.

Read that again. The directive isn’t just about future access. If your SaaS product wrapped Fable 5 last week and you had non-U.S. users, you handled a controlled item. That’s an Export Administration Regulations matter now, and “we already switched to Opus” is not a defense for what happened between June 9 and June 12.

Pro tip: Pull your access logs for June 9-12 before you do anything else. Filter by user country and IP. Save the export. If anyone ever asks who used Fable 5 from your platform, you want that data already sitting in a file, not reconstructed from cold storage six months later.

One more thing the model-string fix misses: the directive covers foreign nationals inside the U.S. too – including Anthropic’s own foreign-born staff. “We’re a U.S. company” is not a safe assumption for your AWS Bedrock deployment if any foreign nationals on your team had access.

The migration paths worth considering

Three real options. None is obviously correct – it depends on what you were using Fable 5 for.

I’d push most teams toward the third option, even though it’s the most work. David Sacks described the restriction as “reluctant” and tied to Anthropic remediating the flagged issue – but no timeline exists. “Reluctant” plus “no date” means you’re running on guesswork if you assume Fable comes back this quarter.

If you were on Mythos 5 specifically – the partner-tier model – Opus 4.8 is not a clean drop-in. Mythos 5 was offered to a smaller contingent of cyber-defenders and infrastructure providers with fewer safeguards than Fable 5. Fable is built on top of Mythos, which is why both got swept into the same directive. Some prompts that worked on Mythos will get refused by Opus 4.8 by design. That’s intentional, but it will break your workflow if you don’t know about it going in.

A concrete dual-route setup

A thin router that prefers Claude Opus 4.8 but falls back to GPT-5.5 if the response gets refused or the API errors out. Buys vendor independence without a full rewrite.

def call_llm(prompt, system=None):
 try:
 resp = anthropic.messages.create(
 model="claude-opus-4-8",
 system=system,
 messages=[{"role": "user", "content": prompt}],
 max_tokens=2048,
 )
 text = resp.content[0].text
 if is_refusal(text): # your own classifier
 raise FallbackNeeded()
 return text
 except (anthropic.APIError, FallbackNeeded):
 resp = openai.chat.completions.create(
 model="gpt-5.5",
 messages=[
 {"role": "system", "content": system or ""},
 {"role": "user", "content": prompt},
 ],
 )
 return resp.choices[0].message.content

The is_refusal check is the part most fallback designs skip. API errors aren’t your main risk anymore – policy refusals are. A model that’s technically up but refuses your prompt is functionally as broken as one that’s offline. Treat both the same way.

What Anthropic actually disputes

Turns out the government’s stated trigger is narrower than the reaction suggests. Anthropic says Commerce provided only verbal evidence of a potential narrow, non-universal jailbreak – essentially asking the model to read a specific codebase and fix any software flaws. A coding task with a specific framing.

If that framing is the real trigger, any model capable of autonomous code repair is theoretically in the same exposure category. Plan accordingly either way – the dual-route pattern above is a hedge against what may now be standard operating procedure across frontier model providers, not just an Anthropic-specific contingency.

FAQ

Can I still use Claude at all from outside the U.S.?

Yes. Only Fable 5 and Mythos 5 were pulled. Opus 4.8, Sonnet 4.6, and Haiku 4.5 are unaffected for everyone, anywhere, as of June 2026.

I run a startup in Europe and my main agent was on Fable 5. What’s the realistic timeline to get Fable back?

Nobody knows, and don’t let anyone tell you otherwise. The administration described the order as reluctant and tied to Anthropic fixing the specific jailbreak – the codebase-repair prompt pattern. That could be a patch in two weeks or a months-long negotiation depending on whether Commerce accepts Anthropic’s fix. If your roadmap depends on Fable returning by a specific date, you’re gambling. Migrate to Opus 4.8 now and treat any return of Fable as a bonus, not a plan.

Do I have to report anything if a foreign user hit Fable 5 through my product before June 12?

That’s a real legal question and worth a 30-minute call with an export-controls attorney rather than an answer from a tutorial. The short version: the directive was issued under the Export Administration Regulations framework – the same legal basis the U.S. uses for hardware export controls – so reporting and recordkeeping obligations are non-trivial. At minimum, preserve your access logs from June 9 through June 12 before anything rotates out of retention. An attorney familiar with BIS (Commerce’s Bureau of Industry and Security) is the right call here, not a blog post.

Next action: open your codebase, grep for claude-fable-5 and claude-mythos-5, and replace them with claude-opus-4-8. Then export your June 9-12 user access logs to a dated file and put it somewhere you won’t lose it. Do those two things before you close this tab.