Claude Mythos 5.1

Claude Mythos 5.1 is a large language model by Anthropic, released September 1, 2026. Like Mythos 5 before it, it ships in two forms: Mythos 5.1, available only through Anthropic’s trusted access programs — including Project Glasswing — and Claude Fable 5.1, which is generally available. Anthropic calls them “the same model, but with different levels of safeguards,” with Mythos 5.1’s “specifically designed to support work in cybersecurity and the life sciences.”¹

The announcement pitches it less as a capability jump than as an answer to customer complaints, naming “price, data retention, and safeguards” as the feedback it addresses. Input and output prices are unchanged; the cost reduction comes entirely from cache reads.¹

Technical details

• Pricing: $10/mtok input, $50/mtok output, identical to Fable 5. Cache reads drop 75% to $0.25/mtok — around 25% cheaper for typical workloads and up to ~45% for context- and tool-heavy agentic ones, by Anthropic’s own indexed measurements over four weeks of August 2026 usage.¹
• Context: 1M tokens in, 128K out. Both the reliable knowledge cutoff and the training data cutoff are June 2026.²
• Reasoning: Adaptive thinking, always on. Effort defaults to high on the API and in Claude Code, and to medium in Claude Cowork and on Claude.ai.²¹
• Data retention: Enterprise Frontier Safeguards (EFS) stores customer data on the customer’s own cloud infrastructure rather than Anthropic’s, which Anthropic describes as “complete privacy (the same as a zero data retention policy)” while still permitting misuse detection. It rolls out in phases from fall 2026; until then, eligible customers get zero data retention.¹
• Safeguards: Retuned for false positives. Anthropic reports Claude Code users can expect around 60% fewer interventions per session from its cyber safeguards than under Fable 5.¹
• API changes: Three breaking — forced tool use now returns an error, thinking blocks are tied to the model that produced them, and editing earlier turns invalidates them. Five additive — per-message effort, turn-scoped system messages, readable progress updates between tool calls, the lower cache-read price, and content provenance.³