Claude Fable 5 Returns With Nationality Controls, Developers Report a More Aggressive Fallback

Fable 5 is back. The outage ran from June 12 to approximately June 18, six days
that forced teams to reroute workloads or pause deployments entirely. Anthropic
restored access, but what returned isn’t quite the same product teams had before.

Two changes are confirmed. First, the restored model ships with nationality-based
access controls and enhanced compliance screening during API onboarding, according
to Microsoft’s Azure
Foundry availability documentation and corroborating TJS registry coverage. Second, the fallback mechanism to Claude Opus 4.8, present before the shutdown
and by design, per Anthropic’s own model documentation, is now triggering more
frequently, at least according to the developers who’ve been using it since
restoration.

The catch is that Anthropic hasn’t published updated fallback rate data. Developers
across Reddit and developer forums report that prompts involving cybersecurity,
chemistry, and biological research are hitting the Opus 4.8 fallback at a rate
that feels meaningfully higher than it did before the shutdown. That’s a
community-observed pattern, not an officially confirmed measurement. Before the
shutdown, Anthropic indicated fallback rates were low, but the specific figure
couldn’t be independently verified from active sources, so don’t treat it as a
hard baseline. What’s confirmed is that the fallback mechanism exists, it’s by
design, and Opus 4.8 is a capable model in its own right. The part nobody mentions
is the cost differential: Fable 5 runs at $10 per million input tokens and $50 per
million output tokens. Opus 4.8, the fallback, runs at $5 input and $25 output. For teams running at scale, unexpected fallback isn’t just a capability question –
it’s a cost question. Verify both price points against the current Anthropic
pricing page before building budget models around them.

The compliance picture has two distinct pieces. Fable 5 is accessible via API with
enhanced screening and nationality-based controls now active. Mythos 5 is not. It
remains offline for general availability; access runs exclusively through
Anthropic’s Project Glasswing program, a restricted channel for government
agencies and approved cybersecurity partners. If your team’s workflow depends on
Mythos 5, that situation hasn’t changed. The Anthropic announcement page
confirms restricted-only access.

On benchmarks: Fable 5 self-reported 80.3% on SWE-Bench Pro and 29.3% on
FrontierCode at launch, per Anthropic. Epoch AI evaluated the pre-suspension model
at launch, those findings carry over as context. No independent evaluation of the
restored version has been confirmed. Treat all benchmark figures as self-reported
until Epoch or a comparable third party publishes restoration-specific results.

What to watch. Two things merit tracking in the coming days. First, whether
Anthropic publishes updated fallback rate data, that number, if released, changes
the calculus for teams whose workloads intersect with the flagged topic categories. Second, the scope of the nationality-based controls. The mechanism is confirmed at
a general level; the specific countries or verification requirements aren’t. Legal
and compliance teams at organizations operating across multiple jurisdictions should
flag this for review before relying on Fable 5 for cross-border deployments.

TJS synthesis

The six-day shutdown produced a model with a tighter operational
envelope than the one it replaced. That’s not inherently bad, Opus 4.8 is
genuinely capable, but teams shouldn’t discover the new fallback behavior in
production. Before restoring Fable 5 to any workflow that touches cybersecurity,
chemistry, or biological research, run a prompt sample against the restored model
and measure actual fallback frequency. That data is more useful than waiting for
Anthropic to publish it.