Overview
Claude Mythos 5.2 is Anthropic's restricted research edition, released on November 9, 2026 on the same underlying weights as Claude Fable 5.2. The distinction sits in access and guardrails: this build lifts some dual-use restrictions and goes only to cybersecurity and life-science organizations cleared through Verification Programs, which puts it out of reach for ordinary paid subscribers.
The most concrete demonstration concerns structural biology. Anthropic had the model predict the full capsid of an adeno-associated virus, the protein shell gene therapy uses to deliver its cargo. The approach was to write code that groups five adjacent chains, run an open folding model on each group, then exploit capsid symmetry to assemble all 60 chains. Twenty-eight of 34 capsids were modeled accurately, and on AAV2 the prediction sits about one atom's width away from the experimental structure.
Anthropic's arrangement is to fence off the high-risk capability while keeping one set of weights: the public half ships with guardrails, the unclamped half goes to organizations equipped to do defense work. That structure has held since 5.1, and 5.2 simply raises the ceiling. For most people Mythos matters as a marker of where the product line reaches, not as something to use.
Key Features
- Gated access: Available only to cybersecurity and life-science organizations approved through Verification Programs, which requires stating the intended use and safeguards. Individual subscriptions and ordinary enterprise accounts cannot call it.
- AAV capsid structure prediction: Groups five adjacent chains, runs each through an open folding model, then uses capsid symmetry to assemble all 60 chains. Twenty-eight of 34 capsids were accurate, and AAV2 landed within roughly one atom's width of the experimental structure.
- Loosened dual-use limits: Compared with the public Fable 5.2, this build eases interception on high-risk biological and cybersecurity requests, aimed at work such as vulnerability defense and biosecurity assessment that needs genuine capability.
- Shared foundation with Fable 5.2: Coding, reasoning and vision gains carry over, along with the 1M context window and output up to 30% faster.
- Institutional billing: Not sold through public plans. Pricing structure and quotas are set by institutional agreement, so individuals cannot buy access separately.
Use Cases
- Gene therapy vector design, predicting whole capsid structures to shorten experimental screening.
- Biosecurity assessment, identifying risks in dual-use research and designing mitigations.
- Vulnerability defense and code auditing within an authorized scope.
- Restricted research pipelines that need the same reasoning power as Fable 5.2.
Pros
- Capsid prediction reaches atomic accuracy, with 28 of 34 structures modeled correctly, which has direct value for vector engineering.
- Sharing a foundation with Fable 5.2 means the public model's coding and reasoning gains carry over.
- Access review routes capability to organizations that can handle it, keeping responsibility boundaries clear.
- A 1M context suits long biological sequences and large codebases.
Pricing
Priced the same as Fable 5.2 at $8 per million input tokens and $40 per million output tokens, with cache reads at $0.20. It is not sold to the public: access requires Verification Programs review, and actual quotas and billing follow the institutional agreement.
Summary
Claude Mythos 5.2 is the restricted edition of the same weights, lifting some dual-use limits for approved cybersecurity and life-science organizations. Its published result is a full AAV capsid prediction, accurate on 28 of 34 structures and within about one atom's width on AAV2. Ordinary users cannot call it; its value is marking the ceiling of this product line.
Version History
- Claude Mythos 5.2 released to vetted organizations only (2026-11-09): Anthropic released Claude Mythos 5.2 on November 9, 2026, sharing underlying weights with Fable 5.2, lifting some dual-use restrictions and limiting access to cybersecurity and life-science organizations cleared through Verification Programs. The headline demo predicts a complete AAV capsid by grouping five adjacent chains through an open folding model and assembling all 60 chains via capsid symmetry: 28 of 34 capsids modeled accurately, with AAV2 within about one atom's width of the experimental structure. Pricing matches Fable 5.2 at $8/$40 per million tokens with cache reads at $0.20.