AI
Anthropic releases Fable 5.1 with cheaper caching, smarter performance, and fewer refusals
The updated model cuts cache pricing by 75% while improving benchmark scores and loosening overly cautious safeguards that were impeding legitimate use cases.
The New Stack
Anthropic unveiled its latest generation of flagship language models on Tuesday, introducing Fable 5.1 and Mythos 5.1 with improvements in capability and economics. The core offering—Fable 5.1—is now publicly available, while Mythos 5.1 remains accessible only through Anthropic's trusted partners, given its specialized safeguards for cybersecurity and life sciences work.
Pricing and availability#
Base token pricing holds steady at $10 and $50 per million input and output tokens respectively, but cache read pricing drops dramatically to $0.25 per million tokens—a 75% reduction. Fable 5.1 continues to require API access and remains available to Max, Team Premium, and Enterprise Premium subscribers (capped at 50% of weekly usage limits), with Pro users able to access it through usage credits.
Performance gains across benchmarks#
Anthropic's own testing demonstrates that Fable 5.1 surpasses its predecessor Fable 5, Opus 5, and OpenAI's GPT-5.6 Sol. The most dramatic jump comes in scientific reasoning tasks: on the Terminal-Bench-Science 0.1 benchmark—which measures agent-based capabilities for research—Fable 5.1 achieved 52.6% versus Fable 5's 24.7%. Other improvements typically cluster in the 2-4% range. The company notes that users can often step down reasoning intensity by one or two levels and still achieve Fable 5 parity, while consuming fewer total tokens. Default reasoning modes now vary by platform: high for Claude Code, medium for Claude Cowork and Claude.ai.
Taming over-cautious guardrails#
A longstanding frustration with Fable 5 was its tendency toward excessive refusals—the model would decline legitimate requests in medical, cybersecurity, and other sensitive domains. Anthropic addressed this by making safeguards more precise: cyber-related protections now trigger roughly 60% less frequently per session, and the model can now identify code vulnerabilities without deferring to older systems. Restrictions remain on penetration testing, exploit creation, and binary scanning. Biology safeguards similarly improved, intervening 85% less on basic medical and biological questions. According to Anthropic, these changes preserve genuine safety while reducing friction on benign queries.
