Today, Anthropic released Claude Fable 5.1 and Mythos 5.1 (the same model with different safety guardrail levels), a mid-cycle minor update with modest performance gains—our primary benefits are improved stability and lower pricing. The official updates are as follows: 1. Greatest improvement in scientific agentic capabilities Terminal-Bench-Science increased from 24.7% in Fable 5 to 52.6% in Fable 5.1—nearly doubling—and significantly surpasses Opus 5’s 29.0%. 2. Enhanced coding and long-running unattended tasks Terminal-Bench 4.0 improved from 42.0% to 55.8% (Mythos 5.1 reached 60.9%); CursorBench 3.2.0 achieved 73.4%. Millennium used it to identify a one-in-a-million crash that had eluded the team for four to five years and was missed by all other models. 3. Prices reduced by approximately 25% Cache read costs dropped 75%, now at $0.25 per million tokens; overall cost for typical workloads decreased by ~25%, and up to ~45% for heavy agentic scenarios. Input remains $10 and output $50 per million tokens. 4. Low-effort settings are sufficient At Low/Medium settings, performance already matches or exceeds Fable 5, at significantly lower cost—ideal for cost-sensitive use cases that can downgrade directly. 5. Drastic reduction in false positives from safety guardrails Cybersecurity guardrails in Claude Code now trigger ~60% fewer blocks, and it is now permitted for vulnerability identification (still prohibited from writing exploits, penetration testing, or binary scanning—these are routed to Opus). False triggers for benign biological queries (basic biology, medical questions) decreased by 85%. 6. Enterprise data retention solution The new Enterprise Frontier Safeguards (EFS) stores data on the customer’s own cloud rather than Anthropic’s, achieving equivalent zero data retention—phased rollout begins this fall; eligible customers may use zero retention immediately until then. 7. Improved alignment over the previous generation Less likely to escape sandboxed environments to access external resources; fewer attempts to justify actions with “this is a simulation/evaluation” reasoning; reduced frequency and success rate of reward hacking; and currently the most robust model against prompt injection. Two additional notes: New API accounts can no longer manually edit Claude’s historical reasoning within multi-turn conversations (to prevent distillation), and outputs include an invisible text watermark required by the EU AI Act—this does not affect content quality or contain any user information.