Overview
Claude Mythos 5 is an Anthropic model positioned for cybersecurity and biology research. It shares underlying model weights with Claude Fable 5, but some safeguards are lifted for authorized uses. Anthropic released it on June 9, 2026. (anthropic.com)
Capabilities and uses
- Accepts text and image input and produces text output. It has a 1 million-token context window and a maximum output of 128K tokens.
- Uses adaptive thinking, which is always on, with a default effort setting of
high. Its stated knowledge and training-data cutoffs are January 2026. - Intended applications include defensive cybersecurity, software engineering, scientific analysis, and life-sciences research. Anthropic reports testing it on protein design and molecular-biology hypotheses. (platform.claude.com)
Access and limitations
Access requires verification; the model is not generally available without approval. Anthropic lists the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry as platforms. The model is active, with retirement no sooner than June 9, 2027. Outputs can be incorrect and should be independently checked, particularly for security or scientific decisions. (platform.claude.com)
The original announcement stated that Mythos-class traffic would be retained for 30 days on first- and third-party surfaces, subject to the policy's stated details and exceptions. (anthropic.com)
Pricing
The following are Anthropic's USD Claude API prices, checked on November 16, 2026. Rates are per 1 million tokens; partner-operated cloud platforms can have independent pricing. The standard rate applies across the full context window. (platform.claude.com)
| Charge | USD per 1 million tokens | Condition |
|---|---|---|
| Input | $10 | Standard, global routing |
| Output | $50 | Standard, global routing |
| Cache write | $12.50 | 5-minute cache |
| Cache write | $20 | 1-hour cache |
| Cache read | $1 | Cache hit |
| Batch input | $5 | Asynchronous Batch API; 50% discount |
| Batch output | $25 | Asynchronous Batch API; 50% discount |
For supported first-party deployments, specifying US-only inference incurs a 1.1× multiplier on token pricing, including input, output and cache charges. Pricing modifiers can stack. (platform.claude.com)