Claude Mythos Preview: Gated Cybersecurity Release
Anthropic announced Claude Mythos Preview on April 7, 2026. This invitation-only research preview supports defensive cybersecurity work under Project Glasswing.
Announcement Details and Scope
The official announcement date for this release is April 7, 2026. On this date, Anthropic stated that Claude Mythos Preview is available as a gated research preview. The documentation explicitly defines the scope of this release as being for defensive cybersecurity work. This initiative operates as part of Project Glasswing.
Access to the model is limited to an invitation-only basis. The source facts do not describe a public application process or general availability timeline. The release notes confirm the provider is Anthropic and the model name is Claude Mythos Preview. No other use cases or broader deployment scenarios are mentioned in the provided evidence.
Documented Specifications and Unknowns
The recorded input modalities for Claude Mythos Preview are text and image. The output modality is listed as text only. The source data does not specify a context window size or a maximum output token limit. Pricing information is also absent from the available records.
The status of open weights is recorded as false, meaning they are not confirmed. This fact does not establish whether local installation is impossible, nor does it mandate API access or define any specific hosting arrangement. Engineers should treat the weight status as an unconfirmed variable rather than a definitive architectural constraint.
- Input Modalities: Text, Image
- Output Modality: Text
- Open Weights: Not Confirmed
- Pricing: Not Specified
Proposed Evaluation Procedure
Researchers with access can evaluate the model by designing tests focused on its documented scope. First, test the model using a standardized set of prompts related to defensive cybersecurity tasks. Measure the relevance and accuracy of the generated text responses against known correct answers or established security standards. Compare these results against baseline metrics from existing internal tools to assess relative performance. Finally, validate consistency by repeating the test suite across multiple sessions to observe stability in behavior.
This procedure allows for factual assessment without inferring operational benefits or architectural advantages. The evaluation remains confined to the defensive cybersecurity domain specified in the announcement. Any conclusions drawn must be based solely on the measured outputs from these proposed tests.
Primary sources
Frequently asked questions
How is access to Claude Mythos Preview granted?
Access is invitation-only. The source states the model is a gated research preview. There is no mention of a public registration method or self-service portal in the provided facts.
Does the source confirm if the model has open weights?
No. The source records that open weights are not confirmed. It does not state whether local installation is possible or if API usage is mandatory. The hosting arrangement remains unspecified.