Claude Fable 5 is back on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork as of today, three weeks after the US government pulled it off the market with an export-control order. The model didn't change. What changed is the layer sitting in front of it — and that layer is now the more interesting product.
Fable 5 shipped on June 9 as the first Mythos-class model Anthropic considered safe for general release: state-of-the-art on nearly every published benchmark, priced at $10 / $50 per million input/output tokens, and roughly half the cost of the Mythos Preview it succeeds. On June 12, the Commerce Department applied export controls to both Fable 5 and Mythos 5, forcing Anthropic to lock foreign nationals out inside and outside the US. Today those controls lift and the model returns — but with a new set of classifiers explicitly targeting cybersecurity misuse.
The safeguard architecture is the story
The public framing around Fable 5's first release was that "some queries will get a response from Opus 4.8 instead." That's a polite way of describing a routing layer: a classifier looks at the incoming request, and if it trips a threshold, the platform silently downgrades the request to a smaller, more constrained model. Anthropic said the classifier fired on under 5% of sessions during the original launch and admitted the tuning was conservative.
Today's redeployment tightens that layer specifically around cyber. The classifiers now catch a broader set of offensive-security tasks — the kind of thing you'd want a state-of-the-art model for if you were building malware, and the kind of thing the export-control conversation was about in the first place.
There are two things worth naming about this design.
First, the classifier stack is now the actual policy surface. The weights haven't moved. Whether Fable 5 will answer your question is a function of what a smaller model decides your question was about. If you care about which questions the frontier can be steered to answer, you should be reading Anthropic's classifier-training methodology, not the model card.
Second, the routing is invisible by default. A user who trips the cyber classifier gets an answer from Opus 4.8 without knowing they were downgraded. That's the right UX for the modal case — most people don't need to know — but it means the model's observed capability distribution is going to look lumpier than the underlying weights, and any external eval that samples across "hard" tasks is going to miss the fact that some of them silently rerouted.
Pricing is the quiet aggression
$10 / $50 per million tokens for a frontier-benchmark-topping model, delivered at Mythos-class latency, is a notable move. It's less than half what Mythos Preview cost, and it undercuts every other model that claims to sit in the same capability tier. Anthropic doesn't usually lead with price, and the fact that they did here suggests two things worth watching:
- They think the model has room on unit economics. Cutting price in half at launch is not what you do if you're still tuning inference cost. It's what you do if the next thing you ship will be even cheaper and you want the market anchored.
- Enterprise routing is where they expect the volume. Priced this way, Fable 5 becomes plausible as a default in agent loops that used to fall back to Sonnet or Haiku for cost reasons. That reshapes the economics of everything downstream of it — including tools like Claude Code, which will now happily burn Fable 5 tokens by default in ways it wouldn't have three weeks ago.
What to actually do about it
For anyone building on the Claude platform:
- Re-benchmark your pipeline. If you evaluated on Opus 4.8 or Mythos Preview in June, run the same suite on
claude-fable-5this week. The pricing shift alone probably changes your model-selection defaults. - Watch for silent reroutes. If you're doing anything adjacent to security, red-teaming, or offensive tooling, expect a non-trivial fraction of your calls to come back from Opus 4.8. Log the response model id; don't assume the one you requested is the one you got.
- Update your data-residency posture. Foreign-national access is back, but the export-control episode won't be the last of its kind. If your product routes model calls per user, now is a good time to make sure you know which jurisdiction each session is anchored to.
The redeployment is a footnote in the launch story. The classifier stack around it is the actual product Anthropic is asking you to trust — and it's the thing worth paying attention to as the next Mythos-class release comes into view.
