Claude Mythos 5.1: verified access and API costs

Mythos 5.1 was released on September 1, 2026. The model guide lists the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Choose a platform covered by your grant and provision the account.

For cybersecurity, apply through Anthropic’s Verification Portal. The October 6 CVP expansion includes Mythos 5.1 across Defense, Red Team and Specialized Access. Independent researchers can apply for Defense Access on a paid plan; Red Team and Specialized Access are for organizations. Choose the tier for work you are authorized to perform.

For life sciences, LSVP is a beta for teams and institutions through the first-party API and Claude Team or Enterprise. Its form registers interest, rather than submitting a formal application. Individual plans and third-party platforms are outside LSVP. Mythos High-risk Use grants are limited to a small set of entities with additional vetting.

After approval, a Claude Console organization owner assigns the program to the intended workspace. Use credentials associated with that workspace and select claude-mythos-5-1 in the Messages API. The example below asks for a bounded code-review plan.

Set a modest output limit for the first request, then inspect the response and usage before adding tool loops. Adaptive thinking is always on and the default effort is high. The output limit caps the response; billing follows actual usage. Cloud-provider provisioning and invoices follow that provider’s route.

For verified CVP users on Pro, Mythos uses usage credits rather than included plan usage. On Max, Mythos and Fable share an allowance of up to 50% of the weekly limit and also count toward the regular limit. Those subscription rules are separate from the API token rates below.

The table uses Anthropic’s Claude API USD rates checked October 12, 2026. For a standard request with 100,000 uncached input tokens and 10,000 billed output tokens, input costs $1 and output costs $0.50: $1.50 total. Repeated tool turns add their own input and output usage.

Prompt caching charges eligible tokens by operation. Writing 100,000 tokens to a five-minute cache costs $1.25; a later read of that prefix costs $0.025. A one-hour write costs $2. Do not charge those same prefix tokens again as ordinary uncached input. Remaining input and each call’s output still add to the bill.

Batch reduces input and output prices by 50% for asynchronous work. The same uncached token volumes would cost $0.75 at the listed Batch rates. US-only inference on the first-party API applies a 1.1× multiplier across token categories, making the standard $1.50 example $1.65. Partner-operated regional endpoints have their own pricing.

LSVP requires 30-day retention for monitoring its traffic. Its data is compartmentalized and excluded from model training and access by Anthropic’s life-sciences research teams. High-risk life-sciences grants apply to a specific project and need renewal every six months; Standard Use grants cover a team and renew annually. Other safeguard domains remain in place.

CVP requires retention by default, but the October 6 update preserves a temporary ZDR exception for organizations already entitled to use Fable 5.1 or Mythos 5.1 with zero retention. Amazon Bedrock CVP requires eligibility for Enterprise Frontier Safeguards (EFS). Provision the applicable agreement and workspace before sending research or security material.

A CVP grant covers internal security work on your own code and infrastructure. A customer-facing product or service using that grant requires separate approval. Account access and the approved use case therefore matter as much as the token price.

Fable 5.1 and Mythos 5.1 share the underlying model, context, output limit and list prices. Their safeguards and access differ. Use Fable for generally available work; consider Mythos when an approved professional task needs the verification program’s scope.

Opus 5.5 lists $4 input and $20 output per million tokens for the Claude API, versus Mythos’s $10 and $50. The standard 100,000-input / 10,000-output example costs $0.60 on Opus and $1.50 on Mythos. Opus also participates in CVP, so verified cybersecurity access does not require choosing Mythos for every task. Compare accepted results and complete tool-loop costs.

Anthropic’s release table reports 60.9% for Mythos 5.1 on Terminal-Bench 4.0, versus 55.8% for Fable 5.1. The publisher explains that the gap reflects safeguards intervening on particular tasks, despite the shared underlying model.

Evaluate a small set of tasks your team can score: a review must identify a real issue, and a proposed patch must pass its relevant checks. Record effort, tool access, input and output usage, retries and review time. Keep the permission scope constant when comparing Fable, Mythos and Opus.