Claude Fable 5 Is Out: Anthropic Opens Up the Mythos That 'Spooked the Government,' With a Gate in Front

On Tuesday afternoon, June 9, Anthropic released Claude Fable 5. The one-line version: the public’s first access to a Mythos-class model.

Mythos needs an introduction. It’s the next-generation system Anthropic began previewing in April, available until now only to partner organizations in a restricted program. NBC’s headline says it’s “built on the same tech that spooked the government.” BBC’s coverage notes Mythos caused a stir among technology, finance, and government leaders. Now Anthropic is letting it out — with a gate.

One model, two products

VentureBeat laid out the structure of the release: Anthropic launched Claude Fable 5 and Claude Mythos 5 at the same time. Same underlying model, different access:

• Fable 5: the public version, available to everyone, safeguards always on
• Mythos 5: the full version, still limited to organizations in the restricted Project Glasswing program

The gate is the most interesting design decision here. AWS’s documentation is specific: a set of classifiers sits in front of the model, covering cybersecurity, biology, chemistry, and health. Flagged requests aren’t answered by Fable 5 — they fall back to Opus 4.8 automatically, the user is told, and rerouted requests aren’t billed at Fable prices.

Answering dangerous questions with a weaker model is smarter than refusing outright: the risk surface stays covered without normal users hitting a wall all day. That said, Anthropic itself admits the launch-day classifiers are tuned “a little too trigger happy” and will be adjusted over time.

The numbers: capability and price both doubled

On capability, the figure that jumps out of VentureBeat’s data is ExploitBench: the model family scores 78.0%, against 40.0% for Opus 4.8 and 34.0% for GPT-5.5. That’s close to a generational gap in high-risk capability, which is exactly why the gate is a requirement rather than decoration. For coding, the widely reported SWE-bench Pro score is about 80.3%, and GitHub announced same-day availability in Copilot, positioning it for “long-horizon, autonomous coding and knowledge-work tasks.”

On price, Sherwood reports Fable 5 costs twice as much as Opus 4.8: $10 per million input tokens and $50 per million output tokens on the API. Compare that with the “same price” posture of the Opus 4.8 launch two weeks ago and Anthropic’s pricing logic is clear: incremental upgrades stay flat, generational capability costs extra.

How access works in practice

AWS’s launch post shows what using Fable 5 involves beyond the price:

• Where it runs: Amazon Bedrock (initially US East and Europe (Stockholm)) and the Claude Platform on AWS, as well as Anthropic’s own API. Bedrock access was rolled out gradually by account.
• Data retention: Anthropic requires 30-day retention of inputs and outputs for Mythos-class models, with human review, so it can spot misuse patterns that are invisible in a single exchange. On Bedrock, customers had to opt in through a data-retention setting before calling the model, and AWS notes that the data then leaves its own security boundary. Teams with strict data-handling rules need to check this before adopting the model.
• Billing on fallback: when a request is routed to Opus 4.8, you pay Opus prices. If a conversation is blocked partway through, earlier tokens are billed at Fable rates and later ones at Opus rates.
• Capabilities Anthropic highlights: long-running asynchronous coding and knowledge work, stronger reading of charts and tables inside PDFs, and checking its own output against the goal.

What happened after launch

The release did not stay smooth. According to editor’s notes on the AWS post, on June 12 Anthropic asked AWS to revoke access to Fable 5 and Mythos 5 for all users to comply with a US government export-control directive; other models, including Opus 4.8, were unaffected. AWS reported access restored on July 1. For anyone building on frontier models, the episode is a reminder that availability can depend on regulation as well as on the vendor, and that a fallback model in your own stack is worth keeping.

The timing is not a coincidence

TechCrunch flagged the awkwardness: days before this release, Anthropic warned that AI is becoming too dangerous. Then it shipped its most powerful public model. Critics will call that safety messaging timed to a product cycle. Anthropic’s counter is that the classifier gate is precisely what makes broad release viable.

Put the other timeline next to it too: on June 1, Anthropic confidentially filed for a U.S. IPO, two weeks after the $965 billion Series H. Funding, IPO filing, flagship release — three moves in 13 days. The pre-IPO story arc is fully assembled.

Our take

The genuinely new thing about Fable 5 isn’t the benchmark scores. It’s tiered capability release as a product shape: one model, where the public gets the gated version and vetted organizations get the full one. If this works, it’s likely to become the standard playbook for how frontier labs ship their strongest models — the first product that gives regulators their accounting and the business its scale at the same time.

Two things to watch short-term. First, the classifier false-positive rate: if the “too trigger happy” launch tuning lingers, developers will vote with their feet. Second, how noticeable the fallback-to-Opus-4.8 gap feels in practice — users are paying for Fable expectations. We’ve started running Fable 5 in our daily workflow, and the hands-on review will be its own article.