← All models

Claude Mythos Preview

Published Publisher Anthropicclaude-mythos-preview

Overview

Claude Mythos Preview is a general-purpose frontier model with advanced agentic coding and reasoning capabilities. Anthropic announced it on April 7, 2026, alongside Project Glasswing, with access restricted to selected partners for defensive cybersecurity work. (anthropic.com)

Capabilities and uses

  • Accepts text and image inputs and produces text output. It can assist with conversation, analysis, coding and creative tasks. (anthropic.com)
  • Intended Glasswing uses include finding and fixing vulnerabilities, binary black-box testing, endpoint security and penetration testing of partner systems. (anthropic.com)
  • Anthropic reports that, after an initial user prompt, the model could autonomously discover and exploit zero-day vulnerabilities in major operating systems and browsers. These capabilities are dual-use and create offensive as well as defensive risks. (anthropic.com)
  • In Anthropic’s evaluations, it scored 77.8% on SWE-bench Pro, 93.9% on SWE-bench Verified and 82.0% on Terminal-Bench 2.0. These are benchmark results under Anthropic’s stated test conditions, not guarantees of real-world performance. (anthropic.com)
  • Anthropic’s pricing documentation describes a 1 million-token context window for this model. (platform.claude.com)

Access and limitations

Anthropic stated that it did not plan to make this preview generally available. The restricted deployment used access controls and real-time classifier safeguards because of cyber misuse risks. As of November 16, 2026, Anthropic’s model lifecycle documentation lists the model as deprecated since June 9, 2026; its retirement date is to be announced. Deprecation does not itself mean retirement. (anthropic.com)

Pricing

Anthropic’s Glasswing announcement stated that, after the research-preview period, participants would be charged USD $25 per million input tokens and USD $125 per million output tokens. It described access through the Claude API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. The announcement also described model-usage credits for the research preview. This is the announced participant pricing, not a verified current public offer or a separate platform-specific price. Anthropic’s current documentation says the full 1 million-token context uses standard per-token pricing; current preview-specific cache and batch amounts have not been independently verified. Source and pricing checked November 16, 2026. (anthropic.com)

Official sources