Anthropic launches Claude Fable 5, but keeps prompts 30 days under new retention rules

Anthropic has shipped Claude Fable 5, a Mythos-class model that the company says is “state-of-the-art on nearly all tested benchmarks” and that it can fail over to Opus 4.8 for certain risky query types. But the part that should jump out to any operator is not the benchmarks. It is the fine print: Anthropic says prompts submitted to, and outputs generated by, Mythos-class models are retained for 30 days for trust and safety purposes on every platform where these models are offered.

Fable 5 is now publicly available as of Tuesday, and Anthropic is also offering Claude Mythos 5 privately for Glasswing partners. Mythos-class models are positioned as a tier above Claude Opus in benchmark performance, and Anthropic frames the naming and safeguards as the key difference between “Fable” and “Mythos.” The company says Fable is capable of causing serious damage without safeguards, which is why it pairs those safeguards with routing. And crucially for procurement and privacy reviews, this new retention policy technically changes what “zero data retention” means for some buyers.

Here is the practical setup. Anthropic says Fable 5 is distinct from “Mythos Preview,” the earlier sibling in the model family. It also says Mythos-class models come with a “new set of classifiers,” separate AI models that look for misuse. Separately, Anthropic describes a failover protocol: for certain types of queries, Fable routes to Opus 4.8. The specifically called out prompt categories are cybersecurity, biology and chemistry, or distillation. In other words, Anthropic is not just turning on safety features. It is designing an operational fallback so the most sensitive requests do not get the full power of the Mythos-class model.

Now add the retention policy, which targets a specific slice of customers. Anthropic says consumer plans (Claude Free, Pro, and Max) are unaffected because Anthropic already has a data retention policy in place for those tiers. The change applies to organizations that have set up workspaces with zero data retention (ZDR) in Claude Console, use Claude Code with ZDR in Claude Enterprise, or access Claude through AWS Bedrock, Google Cloud Agent Platform, or Microsoft Foundry with ZDR. Those customers are told that prompts and outputs for Mythos-class models are retained for 30 days, “on every platform where these models are offered.” The policy is framed as temporary retention “to have a record in the event of misuse,” and Anthropic insists it “won't use this data for training new models.”

For executives, this matters because ZDR contracts and governance programs are often built to guarantee that prompts will not be stored. Anthropic is effectively saying that for Mythos-class models, the operational definition of ZDR has an exception carved out for trust and safety. That second-order effect is bigger than the model launch itself. If your compliance team negotiated ZDR with specific assurances, the procurement and legal workflow now has to account for a 30-day retention window for these specific models, even when ZDR is enabled elsewhere.

The commercial and competitive context is equally sharp. Anthropic lists that two models outperform its own Opus 4.8, OpenAI’s Codex 5.5, and Google’s Gemini 3.1 Pro on the chosen test suite. It does not list Gemini 3.5 Flash [PDF]. On pricing, Fable 5 and Mythos 5 charge $10 per million input tokens and $50 per million output tokens, which Anthropic says is less than half the price of Claude Mythos Preview ($25/m input, $125/m output). Anthropic also claims Fable 5 is more token efficient than predecessors. That is relevant because higher capability often comes with higher cost, and executives always end up asking whether the unit economics pencil out once the model becomes the default workflow.

There is also a proof point, and it reads like a CIO-friendly ROI story. Anthropic claims Stripe used Fable 5 to migrate its 50-million-line Ruby codebase in one day, while Anthropic says it would have taken a Stripe team two months without AI help. Whether you care about the specific project or not, the point is that Anthropic is leaning into software engineering as a flagship use case for this tier. Fable 5 is also said to deliver stronger resistance to prompt injection attacks than previous offerings, with a model safety card claiming a 4.8 percent success rate over 100 attempts, comparable to Claude Opus 4.8.

Finally, there is a rollout and licensing cliff. As of today (through June 22), Fable 5 is available through Pro, Max, Team, and seat-based Enterprise plans at no extra cost. After June 22, Anthropic says Fable 5 will no longer be available through those plans and will require usage credits. It also signals that it may restore access in subscription plans later “if capacity allows.” For boards and senior leadership teams, that is not just a product change. It is a capacity and margin decision that will affect adoption timelines, internal budgeting, and how quickly teams can standardize on Mythos-class models without surprise spend.

The strategic stake is clear: Mythos-class models are now entering production-like workflows, and Anthropic is trying to keep them usable without turning them into governance nightmares. But the new 30-day retention requirement for some ZDR setups is the kind of policy shift that can derail rollouts if it is discovered too late. If you are evaluating model providers, the lesson is to treat safety features, routing behavior, and data retention terms as a single procurement package, not separate checkboxes.