Anthropic’s Claude Fable 5: Mythos-class AI with guardrails

On June 9, 2026, Anthropic launched Claude Fable 5, the first Mythos-class model aimed at general use, and Claude Mythos 5 for a smaller trusted circle of cyber defenders and infrastructure partners.

Official announcement: Claude Fable 5 and Claude Mythos 5.

If you only remember one framing: Fable and Mythos share the same underlying model. Fable is the version with safety classifiers attached so Anthropic will sell it broadly. Mythos lifts some of those safeguards for people Anthropic trusts with dual-use capability.

What Fable 5 is supposed to be good at

Anthropic positions Fable 5 above the Opus line on software engineering, long-horizon agent work, knowledge work, vision, and research-style tasks. Early customer quotes in the launch post lean hard on coding agents, finance analysis, and multi-hour autonomous runs.

For MSP and ops work, the useful claim is not a leaderboard screenshot. It is whether the model can hold a messy repo, a long runbook, and a multi-step change without losing the plot, and whether you can afford the tokens when it does.

Guardrails are the product, not a footnote

Mythos-class models are strong enough in cybersecurity and biology that Anthropic will not ship the raw capability to everyone. On Fable 5, separate safety classifiers watch for cybersecurity, biology/chemistry, and model-distillation topics. When they trip, the session falls back to Claude Opus 4.8, and the user is told that happened.

Anthropic says those fallbacks hit fewer than 5% of sessions on average, and that the classifiers are tuned conservatively on purpose (expect some false positives on benign security talk). If your day job is vulnerability research or red-team tooling, Fable may feel nerfed in exactly the moments you care about. That is intentional.

Mythos 5 is the same brain with some safeguards lifted, initially via Project Glasswing and related trusted-access paths. It is not a public “turn off the filter” toggle.

Pricing and access (launch day)

At launch, Anthropic listed Fable 5 and Mythos 5 API pricing at $10 per million input tokens and $50 per million output tokens, less than half of Claude Mythos Preview pricing, still a premium versus Opus-class models.

Subscription rollouts were staged (included usage for a short window, then credits / capacity-dependent). API and consumption Enterprise plans got full availability sooner. Check current Claude console docs for what your plan includes today; launch-week entitlements change fast.

API model id to look for: claude-fable-5.

Brief timeline note (post-launch)

A few days after launch (June 12), Anthropic suspended Fable 5 and Mythos 5 access amid safeguard/export-control pressure, then restored them globally on July 1, 2026. If you are reading older screenshots that say “unavailable,” that gap is why. For day-to-day planning, treat Fable as a generally available Mythos-class model with classifiers, and keep Opus (or another fallback) in your routing policy for when the safety layer intervenes.

How we would actually use it

Do:

• Use Fable for long coding/agent sessions where Opus was the ceiling last month.
• Keep prompts free of production secrets; same rules as any cloud LLM.
• Expect occasional “this reply came from Opus 4.8” moments on security-adjacent questions and design around that.

Don’t:

• Assume Fable will walk you through offensive exploit development the way an unrestricted research model might.
• Point every chat at Fable without watching token cost. Premium models punish sloppy context dumping.
• Skip human review on anything that touches tenants, firewalls, or identity. Stronger models still hallucinate confidently.

Bottom line

Fable 5 is Anthropic’s bid to put Mythos-level capability in front of normal users without shipping unrestricted cyber/bio uplift. Same core model as Mythos 5; different access and safeguards. For practical work, try it where long-horizon coding and analysis matter, and keep a cheaper, less-restricted model in the toolbox for the sessions the classifiers bounce.