Claude Mythos 5.1: Access Is the Product Constraint

If you opened this page looking for a public API key or a pip install path for Mythos 5.1, stop there. Anthropic describes Mythos 5.1 as available only through trusted-access programs for vetted cyberdefenders and life scientists. It is currently available only to a set of US organizations, while Anthropic says it is coordinating with the US government to broaden access.

What Mythos is, and how it relates to Fable

Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguard levels. Fable is broadly available through Anthropic’s standard distribution. Mythos is offered only to approved customers in Project Glasswing. Public access material does not document a self-serve Mythos API identifier or endpoint contract, so treat its interface as program-specific even though the underlying model is shared.

Exactly who can access Mythos today

Per the release page and FAQ, Mythos 5.1 is available to vetted cyberdefenders and life scientists through trusted-access programs. Anthropic states Mythos is currently available to a set of US organizations and is coordinating to broaden access. There is no public self-serve sign-up, consumer tier, or documented general-purpose API listing. This is not a product to buy like a standard API plan.

What to prepare before you contact Anthropic

Anthropic’s public material does not publish CVP application criteria or a life-science intake checklist. The CVP portal is gated. Do not treat an online checklist as official requirements. Before you contact Anthropic, write a short internal brief that identifies your organization, the defensive or scientific work you want evaluated, the data involved, the human-review boundary, and the person accountable for the engagement. That is prudent preparation, not a published eligibility test.

The CVP route, link-only

Cyberdefenders can register interest at Anthropic’s Cyber Verification Program portal. Anthropic’s FAQ says CVP will include Mythos access in the near future; verify with an Anthropic account team whether access is currently provisioned through CVP or a separate Project Glasswing invitation. Anthropic’s public material does not publish a comparable self-serve life-science application page.

Evaluation plan if admitted

Treat the access tier as a constraint, not a generic API. Do not assume a public model identifier, rate limit, or endpoint shape. Ask your Anthropic contact for the authorized interface and terms for your program. Anthropic’s public prompt-caching documentation lists the same published rates for Mythos 5.1 as Fable 5.1: $10 per MTok input, $50 per MTok output, and $0.25 per MTok cache reads. Those rates are a starting point; your program agreement governs actual billing and any commitment terms.

The Fable FAQ says that, for most Claude applications, flagged cybersecurity requests route to Opus 4.8 and biology requests to Opus 5. That is Fable behavior, not a published Mythos routing contract. For Mythos, ask Anthropic to document the routing and safeguards that apply to your program. Treat safeguards as part of the evaluation, not a bug to bypass. Pick a representative, authorized slice of work and instrument it for latency, refusal patterns, citation behavior, and human-review interventions.

Data and safeguard boundaries

Fable 5.1 requires 30-day data retention unless Anthropic expressly authorizes another arrangement. Anthropic’s public material does not state that its Enterprise Frontier Safeguards or zero-retention terms automatically apply to Mythos. Confirm retention, data handling, and routing in your program agreement before designing experiments around them. Use the System Card for safety scope, not as operational documentation.

Vendor benchmark caveats

Anthropic reports different benchmark results for Fable and Mythos where safeguards intervene. Read those as vendor-published results under stated evaluation conditions, not proof that your organization will see the same result. The safeguard layer changes what users can do at inference time, especially on cyber and biology prompts. Build an authorized workload that measures capability alongside refusal correctness, over-refusal on benign inputs, routing behavior, and human-review interventions.

Failure modes to plan for

• Signing up as if Mythos were a public API. No retail plan, no console toggle.
• Inferring a Mythos model ID from Fable’s. The two surfaces are not documented as equivalent.
• Treating listed token rates as a complete contract. Pricing is published, but your program agreement governs access and billing terms.
• Treating EFS or zero retention as default for Mythos. Retention and routing are program-specific.
• Trying to bypass safeguards via prompt craft. The safeguards are what you are evaluating.
• Pulling weights locally. Mythos is not a local model — no documented local-serving path.
• Treating the System Card as exploitable documentation. Use it for safety scope.

Done means

You have a sourced picture of Mythos’s limited availability, the Project Glasswing and CVP references, its published list rates, and what you should not assume about model IDs, endpoints, retention, or routing. If you are a cyberdefender, you know the CVP interest route and will verify its current Mythos provisioning status. If admitted, your evaluation plan asks for documented terms before it touches production data. You have not invented a model ID, endpoint, or retention setting.

What this guide does NOT cover

It does not describe undocumented Mythos endpoints, model identifiers, API shapes, or program-specific billing terms. It does not reproduce parts of the System Card that describe exploitable capabilities. It does not provide ways to bypass safeguards, generate exploits, or run Mythos weights locally. It does not cover the Fable 5.1 general-purpose API surface, which is a separate product.

Related guides

• Claude Opus 5 model profile
• Claude Sonnet vs Haiku vs Opus: when to pick which
• Hermes Agent — Anthropic API provider setup

Sources

• Anthropic — Claude Fable and Mythos 5.1 release
• Anthropic — Claude Fable product FAQ
• Anthropic — Claude Vulnerability Protection Program (CVP) application portal
• Anthropic — Claude Fable 5.1 / Mythos 5.1 System Card
• Anthropic — Prompt caching
• [Claude Fable 5.1 and Claude Mythos 5.1 research dossier](Johnny5/Websites/Agentic Botsitting/research/claude-fable-mythos-5-1/claude-fable-mythos-5-1-research-2026-09-01.md)