Anthropic Launches Claude Fable 5 and Mythos 5: Mythos-Class Capability Opens Up, With Classifier Fallbacks and 30-Day Retention

On June 9, 2026, Anthropic launched Claude Fable 5 (API model ID claude-fable-5) and Claude Mythos 5 (claude-mythos-5). Anthropic describes Fable 5 as a Mythos-class model made safe for general use; Mythos 5 is the same underlying model with safeguards lifted in some areas, initially deployed only to cyber defenders and infrastructure providers through Project Glasswing, in collaboration with the US government. Both are priced at $10 per million input tokens and $50 per million output tokens. According to the Claude Platform release notes, both support a 1M token context window by default, 128k max output tokens, and always-on adaptive thinking.

What changes most is the terms of use. Fable 5 ships with new safety classifiers: when a request touches cybersecurity, biology and chemistry, or model distillation, the response is handled by Claude Opus 4.8 instead, and the user is told. Anthropic says more than 95% of Fable sessions involve no fallback. At the API level, declined requests return stop_reason "refusal", and a beta fallbacks parameter can re-run refused requests on another model at that model's rates. Anthropic also requires 30-day retention for all Mythos-class traffic on first- and third-party surfaces; Fable 5 is not available under zero data retention (ZDR).

There are breaking API changes too. On Fable 5 and Mythos 5, adaptive thinking is the only thinking mode: disabling thinking, manual extended-thinking budgets, and assistant prefill return a 400 error. Both use the tokenizer introduced with Claude Opus 4.7, which produces roughly 30% more tokens for the same text. Upgrading is therefore not just a model-name swap; cost estimates and request parameters need to be re-validated.

Benchmark leadership, the Stripe codebase-migration anecdote, and quotes from Cursor, GitHub, and other early-access customers are vendor claims in Anthropic's own post, not independent results. The statement that the classifiers are tuned conservatively and will sometimes block harmless requests is also Anthropic's own acknowledged tradeoff; the real false-positive rate depends on each organization's workload.

Before adopting, enterprises can work through a checklist drawn from Anthropic's documentation. First, confirm data governance allows 30-day retention and that workloads requiring zero data retention are kept off Fable 5. Second, make sure applications handle stop_reason "refusal" and decide whether to enable the beta fallbacks parameter (not supported on the Message Batches API), knowing fallbacks are billed at the fallback model's rates. Third, remove settings that disable thinking, set manual extended-thinking budgets, or use assistant prefill, all of which return 400. Fourth, re-measure prompts with the token counting API using claude-fable-5, since the newer tokenizer produces roughly 30% more tokens. Fifth, if readable reasoning summaries are needed, set thinking.display to "summarized", because the default is "omitted".

Three things to watch: whether classifier false positives fall as promised, whether the Mythos 5 trusted access program expands beyond Glasswing partners, and whether the 30-day retention rule extends to more future models. Together these determine whether regulated industries can put Fable-class models into production workflows.