Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Fable 5 is the version general users can touch: a Mythos-class model with safeguards that route some sensitive requests to Claude Opus 4.8. Mythos 5 is the same underlying model with some safeguards lifted, but access starts with a small trusted-access group rather than the public product.
That split is the story. Anthropic is not simply shipping a stronger Claude tier. It is turning one model family into two products: a broad model with a fallback path for high-risk domains, and a restricted model for cyberdefenders, infrastructure providers, and other vetted users.
The product split matters
Fable 5 exists because Anthropic wants to put Mythos-level capability into general use without releasing the full model boundary to everyone. The company says the model is stronger on software engineering, knowledge work, vision, scientific research, and long complex tasks. But the release also says that without safeguards, some capabilities could be misused, especially in cybersecurity.
The safeguard design is direct: when a request falls into certain sensitive areas, Fable 5 can hand the response to Claude Opus 4.8. Anthropic names cybersecurity, biology, chemistry, and model distillation as areas where routing can happen. Users should read that carefully. Fable 5 is not always Fable 5. In flagged cases, the user may get Anthropic’s next-most-capable model instead.
That is a product compromise, not just a safety footnote. If your workflow is ordinary software work, writing, research, or visual reasoning, Anthropic says most sessions should stay on Fable. If your workflow sits near security research, dual-use biology, chemistry, or model extraction, the model may behave like a more conservative front end.
Mythos 5 stays behind the gate
Anthropic is also releasing Claude Mythos 5, but not as a normal public model. The company describes it as the same underlying model as Fable 5 with some safeguards lifted. Initial access is aimed at a small group of cyberdefenders and infrastructure providers, with a broader trusted-access program planned over time.
That keeps Project Glasswing at the center of Anthropic’s Mythos strategy. The company has spent the past few months treating Mythos-class cyber capability as unusually sensitive: useful for defenders, risky if handed to anyone asking for exploit help. Fable 5 is the general-access answer to that tension. Mythos 5 is the controlled-access answer.
The result is a clearer ladder than Anthropic had before. Opus 4.8 remains the broadly capable fallback model. Fable 5 becomes the high-capability public model with routing safeguards. Mythos 5 becomes the restricted model for users Anthropic is willing to trust with fewer restrictions.
The 95% number is the testable claim
Anthropic’s useful claim is not just that Fable 5 is safer. It is that the fallback system should rarely interrupt ordinary use. The company says that in testing, 95% of Fable sessions ran without falling back to Opus 4.8.
That figure is worth testing in real workflows. A developer using Fable 5 for everyday code generation may never notice the routing layer. A security team using it for defensive research may hit the boundary often, even when the intent is legitimate. Anthropic says the safeguards are conservative at launch, so harmless requests can be routed away from Fable if they look too close to a restricted area.
The price confirms the tier
Fable 5 and Mythos 5 are priced above Opus 4.8 at $10 per million input tokens and $50 per million output tokens. That makes them twice the standard Opus 4.8 rates, but below the earlier Mythos Preview pricing reported around the restricted program.
That price point tells teams how Anthropic sees the tier: not a replacement for every Claude call, but a premium model for work where the extra capability pays for itself. Long coding tasks, complicated analysis, visual reasoning, and scientific workflows are the places to test first. High-volume extraction, summarization, and routine chat probably stay on cheaper models unless Fable’s quality delta is obvious.
The operational question is simple: does Fable 5 finish work that Opus 4.8 cannot, or does it merely produce a better first draft at twice the token price? That answer will be workload-specific.
What teams should test
For software teams, start with long tasks that currently break across multiple Claude turns: multi-file refactors, bug hunts, architecture reviews, and codebase questions that require maintaining context. Anthropic says Fable’s lead grows on longer and more complex work, so short prompts are the wrong benchmark.
For security teams, record routing behavior explicitly. If benign defensive tasks fall back too often, Fable 5 may be less useful than its headline capability suggests. If Mythos 5 access is available through a trusted program, compare the same tasks across Fable, Mythos, and Opus 4.8 with clear logs of which model answered.
For buyers, separate three questions: performance, routing, and price. A model can be stronger and still be the wrong default if safeguards interrupt the task or if the cost only pays off on a small slice of work.
For broader context on model capability and pricing, see our AI model leaderboard. For Anthropic’s funding, product, and model coverage, see our Anthropic company tracker.
