Claude Fable 5 suspended by US government order, then restored on July 1: what happened and what it means

Update — July 1, 2026: Fable 5 is back. The US Department of Commerce lifted the export controls and Anthropic restored Claude Fable 5 and Mythos 5 on July 1. Access returns behind a new safety classifier aimed at the specific reported bypass — Anthropic says it blocks that technique “in over 99% of cases,” redirecting those requests to Opus 4.8. What it means in practice: Fable 5 is included in Pro, Max, and Team plans at up to 50% of your weekly limits through July 7, then runs on usage credits; cloud-provider access (AWS, Google Cloud, Microsoft) is being restored separately; and Mythos 5 is back only for approved US organizations. The claude-fable-5 API answers again, so the hard 404 described below is now history. The rest of this page stands as the record of the nearly three weeks it was gone — and why the precedent still matters even though the order was reversed.

On the evening of June 12, 2026, Anthropic disabled Claude Fable 5 and Claude Mythos 5 for every customer to comply with a US government export-control directive. It is, as far as anyone has documented, the first time a leading AI company has taken a publicly deployed model offline because the federal government ordered it to. Fable 5 had been on sale for three days.

This page is the developer-facing version of the story: the exact timeline, what your API calls and Claude Code sessions did during the outage, the jailbreak that triggered it, the five-minute migration people ran, and what changed when access came back on July 1. The primary sources are at the end.

Update — June 13, 2026: Multiple outlets now report what set this off. Per the Wall Street Journal, Amazon CEO Andy Jassy told Treasury Secretary Scott Bessent and other officials that Amazon researchers had used Fable 5 to obtain information usable in cyberattacks; the export-control directive followed (TechCrunch, The Information). Amazon told the WSJ that being asked for “counsel on potential security risks” is “not uncommon” and that it doesn’t disclose details. The reporting ties the “national security” rationale to a specific source — a rival’s CEO, and one of Anthropic’s own largest investors — without changing what developers should do: the model is still offline, and the migration below stands.

The 30-second version

1. What happened: a US export-control directive ordered Anthropic to cut off Fable 5 and Mythos 5 for foreign nationals; to comply fully, Anthropic disabled both for all customers. Every other Claude model is unaffected.
2. The API, during the outage: claude-fable-5 returned a hard 404 — "Claude Fable 5 is not available. Please use Opus 4.8." with no auto-redirect. As of July 1 it answers again (see the update above).
3. What you did: pin Opus 4.8 (/model opus, or claude-opus-4-8 in your harness) — a five-minute change at half the price. You can re-pin Fable now, mindful of the July 7 credit cutover.
4. The twist: the flagged “jailbreak” was asking the model to read a codebase and fix its flaws — ordinary agent work, and a capability Anthropic and outside researchers note is “widely available from other models.”
5. Did it come back? Yes — access was restored on July 1, 2026 after the Commerce Department lifted the controls.

The timeline

All times June 12, 2026:

• 5:21pm ET — Anthropic receives the directive. In its words, the order is to suspend access “by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.”
• ~9:58pm ET — last successful Fable 5 API request logged (per Simon Willison’s monitoring).
• ~9:59pm ET / 6:59pm PT — API access cut off. Requests begin returning 404.

The gap matters for one practical reason: a directive received at 5:21pm took effect the same night. If you were running an overnight agent loop on Fable, it stopped mid-run with a 404 it almost certainly didn’t know how to handle.

What your API calls did during the outage

Restored July 1 — this is now history, but worth keeping as a defensive pattern. During June 12–30, a request to claude-fable-5 returned:

HTTP 404
"Claude Fable 5 is not available. Please use Opus 4.8.
 Learn more: https://www.anthropic.com/news/fable-mythos-access"

There was no automatic redirect. This was the failure mode that bit: if you read our fallback-handling guide, you’ll recognize that Fable used to refuse by returning an HTTP 200 with stop_reason: "refusal". During the outage that became a transport-level 404 instead — code that only branched on stop_reason never saw it; code that branched on status codes did. The minimal fix at the time:

# Before: model="claude-fable-5"
# After:
response = client.messages.create(model="claude-opus-4-8", ...)

If you built the refusal branch from our fallback guide, it already pointed at Opus 4.8 — but it triggered on stop_reason == "refusal", not on a 404. The status check that covered both:

try:
    resp = client.messages.create(model="claude-fable-5", ...)
except APIStatusError as e:
    if e.status_code == 404:          # Fable suspended
        resp = client.messages.create(model="claude-opus-4-8", ...)

What Claude Code users should do

With Fable 5 restored, this is back to a normal buy-or-skip decision — the one the verdict page covers in full. Two things worth knowing:

• Opus 4.8 held up as the default, and still does. During the outage the move was to pin Opus (/model opus, or claude-opus-4-8). It cost little: per Anthropic’s own docs, Fable fell back to Opus 4.8 “often on the first request” for penetration testing, CTF, and biology-adjacent work, so for a large slice of agent tasks the model actually answering was already Opus — at half the price ($5/$25 vs Fable’s $10/$50 per million tokens).
• Mind the July 7 cutover. Through July 7, Fable 5 is included in Pro, Max, and Team plans at up to 50% of your weekly limits; after that it draws on usage credits. If you re-pin Fable for the hard, single-shot tasks where its edge is real, watch that your weekly quota and credit balance behave the way you expect.

ANTHROPIC_DEFAULT_FABLE_MODEL points at a live model again. If you added a 404-handling branch or routed Fable to Opus during the outage, leave it in as a defensive pattern — the new classifier can still redirect the flagged read-a-codebase-and-fix-it technique to Opus 4.8.

The jailbreak that triggered it

Here is the part that should give this audience pause. Anthropic’s description of the government’s concern:

the government believes it has become aware of a method of bypassing, or “jailbreaking” Fable 5 — asking the model to read a specific codebase and fix any software flaws.

That is not an exotic exploit. That is the single most common thing people use a coding agent for. “Read this repo and fix the bugs” is the core loop of Claude Code, Codex, Cursor, and every tool we cover.

Anthropic says it received only “verbal evidence of a potential narrow, non-universal jailbreak” — nothing in writing — and disagrees “that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.” Simon Willison, who tracked the cutoff live, makes the capability point bluntly: the demonstrated ability is

“widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe.”

In other words: if reading code and proposing fixes is the dangerous capability, it’s not a Fable 5 capability — it’s a 2026 coding model capability, and the same people who’d be “harmed” by it are the security engineers who rely on it daily.

Why this matters beyond one model

Three things are new here, regardless of where the jailbreak claim lands:

• Precedent. A deployed, commercially available frontier model was pulled by government directive, overnight, affecting paying customers worldwide. Export controls have shaped who can buy chips and models before; this is the first time one reached in and switched off a running product.
• The blast radius of “comply fully.” The order targeted foreign nationals. Anthropic’s read was that the only way to guarantee compliance was to disable the models for everyone — so a directive aimed at a subset took the product down for all of it. That’s a pattern worth understanding if you build on any model that could attract similar attention.
• Single-vendor risk is now also single-jurisdiction risk. If a model can vanish on a few hours’ notice for reasons its own maker disputes, “which model” is no longer only a quality-and-price decision. Keeping a working fallback to a second model — and a second vendor — stopped being a nice-to-have on June 12.

How it came back

It came back on July 1, 2026. After nearly three weeks of talks between Anthropic and the Commerce Department, the government lifted the export controls, and Anthropic redeployed Fable 5 and Mythos 5 the same day behind a new safety classifier — one that, per Anthropic, blocks the specific reported bypass “in over 99% of cases” by redirecting those requests to Opus 4.8. Alongside it, Anthropic proposed a framework for scoring jailbreaks on four axes — capability gain, breadth of that gain, ease of weaponization, and discoverability — which it says it is drafting “together with Amazon, Microsoft, Google, and other Glasswing partners,” plus a new HackerOne program for researchers to submit cyber jailbreaks they find in Fable 5. Amazon co-authoring the rubric is its own irony: per the WSJ reporting above, it was Amazon’s CEO who reportedly raised the concern that pulled Fable in the first place. The restoration carries conditions — 50% of weekly limits through July 7 and usage credits after, with cloud providers and Mythos 5 on separate tracks — but the core outcome is simple: the model is back.

What we watched during the outage (June 18, 2026 snapshot): Six days in, access was still off and Anthropic’s statement page carried no restoration notice. Reporting described in-person talks in Washington as ongoing; restoration rumors ran well ahead of facts (a June 16 X post claimed a “48-hour” return, and a prediction market plus an “is it back yet” tracker sprang up). We flagged only two signals as trustworthy — a notice on Anthropic’s own page, and the claude-fable-5 API returning something other than a 404. Both flipped on July 1.

What would change this page

The verdict page now carries the buy-or-skip framing again. What would still change this page: a re-suspension or a new directive; the new classifier proving over- or under-aggressive in real use (tell us what you see); or the same order reaching other models.

Companion reading

• Claude Fable 5 vs Opus 4.8: the verdict — what Fable is, the real coding gains, and the price math, now that it’s back
• Detecting Fable 5 fallbacks in agent code — the refusal-branch guide, more relevant now: the new classifier redirects the flagged technique to Opus 4.8
• Claude Code usage limits, explained — pinning Opus 4.8 and what it costs against your quota
• Claude Code’s “China backdoor,” explained — another “the tool changed under you” story, this time a covert vendor experiment

Sources

1. Anthropic — Statement on the US government directive to suspend access to Fable 5 and Mythos 5
2. Simon Willison — live tracking, timeline, and analysis (June 13, 2026)
3. Bloomberg — Anthropic says US limits foreign access to Fable 5, Mythos 5
4. CNBC — Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive
5. NBC News — Anthropic suspends new AI models after government directive
6. Anthropic — Redeploying Claude Fable 5 (June 30, 2026)

FAQ

Why was Claude Fable 5 suspended? A US government export-control directive on June 12, 2026 (5:21pm ET) ordered Anthropic to suspend Fable 5 and Mythos 5 access for foreign nationals, citing a national-security concern over a reported jailbreak. To comply fully, Anthropic disabled both models for every customer. Anthropic publicly disagreed, saying it received only verbal evidence of a narrow, non-universal jailbreak.

What happens if I call the Fable 5 API now? As of July 1, 2026 it responds normally again — the export controls were lifted and Anthropic restored access. During the June 12–30 outage claude-fable-5 returned an HTTP 404 (“Claude Fable 5 is not available. Please use Opus 4.8.”) with no automatic redirect. Worth still handling: the new classifier redirects the flagged read-a-codebase-and-fix-it technique to Opus 4.8 in over 99% of cases.

Is Claude Fable 5 back? Yes — restored July 1, 2026 after the Commerce Department lifted the export controls. It returns behind a new safety classifier (blocks the reported bypass in over 99% of cases, redirecting to Opus 4.8). Through July 7 it’s included in Pro, Max, and Team plans at up to 50% of weekly limits; after that it runs on usage credits. Cloud providers and Mythos 5 are on separate restoration tracks.

What should Claude Code users do? Pin Opus 4.8 (/model opus). Fable was already falling back to Opus on security and biology work, so for most agent tasks the practical capability you keep is close — at half the price.