Anthropic says its new Claude Fable 5.1 will cut typical token-billed costs by about 25%, with agent-heavy workflows seeing up to roughly 45% savings, while a stricter sibling, Claude Mythos 5.1, arrives under trusted access for cybersecurity and life sciences work (Anthropic). The company is also introducing Enterprise Frontier Safeguards (EFS), a system that keeps customer data in infrastructure the customer controls, promising the effect of zero data retention alongside stronger misuse prevention.
What Anthropic is changing with Claude Fable 5.1
According to Anthropic, Claude Fable 5.1 and Claude Mythos 5.1 are the same underlying model configured with different safeguards (Anthropic). Fable 5.1 is broadly available. Mythos 5.1 is gated through trusted programs, with controls tailored for work in cybersecurity and the life sciences.
The headline change for buyers is pricing. Anthropic says the effective 25% reduction for common usage comes from lowering the cost of cache reads—when the model reuses previously processed inputs—which matters most in longer, stateful sessions. For highly agentic tasks, where assistants loop through tools, context, and plans, the savings can reach about 45% because large portions of the context are read from cache rather than reprocessed each step.
On privacy, Anthropic is rolling out EFS, which it describes as storing data in cloud infrastructure fully controlled by the customer. The company frames this as offering the same outcome as a zero data retention policy while maintaining strong defenses against adversarial use. Until EFS is in place for a given customer, Anthropic says eligible users can run Fable 5.1 with zero data retention.
Safeguards have been tuned to reduce false positives that block benign content. In cybersecurity, Anthropic reports its newest safeguards cut false positives by 60%. The firm also states the model can assist with discovering software vulnerabilities but is restricted from generating exploits, reflecting a separation that aligns with common vulnerability disclosure norms (CISA VDP guidance).
Pricing math: where agentic work feels the drop
The pricing change targets a real pain point for teams building long-running assistants and RAG pipelines. In many agent setups, the system rereads large context windows—tools lists, instructions, retrieved docs—at each step. If much of that is billed as lower-cost cache reads, effective spend falls without code changes. That’s why Anthropic’s own figures show stronger gains on agentic workloads.
The impact scales with how much of a session’s context is repeatedly reused. Consider a help-desk agent that cycles through retrieval, summarization, and action planning. A high share of the prompt is stable across turns. Lower-cost cache reads can pull total spend down meaningfully across a full workday. Teams running retrieval-augmented generation will see a similar pattern when large chunks of retrieved knowledge stay constant across follow-up questions (IBM on RAG).
Two practical consequences follow. First, engineering time aimed at shrinking prompts may yield diminishing returns if cache-read pricing already covers the stable parts cheaply. Second, product managers can revisit throttles on tool use and context refresh frequency. If the budget’s main risk was context bloat, the new pricing may let teams increase depth—more steps per task, more checks, more citations—without blowing past monthly caps.
EFS and zero data retention: enterprise control without a lab buildout
EFS is the other big lever for regulated buyers. Anthropic says enterprise data are kept in infrastructure under customer control, delivering the practical effect of zero data retention while the provider maintains strong misuse defenses (Anthropic). For security leaders, this mirrors a well-known pattern in cloud security—customer-controlled storage and keys reduce exposure and simplify discovery obligations (AWS KMS basics).
This design can shorten procurement cycles. Data residency, retention, and breach fallout are common blockers in AI pilots. If sensitive data never live in a vendor-controlled environment, legal and compliance reviews often move faster. That maps cleanly to the governance and data-management priorities many organizations follow under frameworks like NIST’s AI Risk Management Framework (NIST AI RMF).
What to verify in a proof of concept: how EFS integrates with your existing identity and access controls; who can see logs and prompts; how encryption and key rotation are handled; and how zero-retention settings are enforced for backup, observability, and fine-tuning pipelines. Buyers should also confirm that EFS performance overhead does not erase the pricing gains promised by cache reads.
Mythos 5.1 and dual‑use: guardrails for cyber and bio work
Anthropic is drawing a line between general availability and dual‑use domains. Claude Mythos 5.1 is restricted to trusted access programs. The company says Mythos safeguards are engineered to support real cybersecurity and life-science tasks while preventing misuse, and it highlights fewer benign blocks as a key upgrade. Anthropic also states the model can assist in discovering software vulnerabilities but will not help develop exploits, which mirrors how security teams separate finding from weaponizing flaws (CISA guidance).
For security leaders, the mix matters. Fewer false positives reduce friction in red/blue team workflows and lower the odds that playbooks stall on harmless content. Allowing vulnerability discovery while blocking exploit creation speaks to maturing, domain‑specific policies rather than one-size-fits-all filters. In bio, a vetted-access model paired with clear constraints signals closer coordination with public-sector partners, which Anthropic says includes a program developed with the U.S. government (Anthropic).
What to watch next with Claude Fable 5.1
Enterprises now have a straightforward checklist. Reprice agent workloads under the new cache-read terms. Pilot EFS with a narrow, high-sensitivity use case and pressure-test auditability. For teams in cyber and bio, request details on Mythos eligibility and escalation paths when safeguards trigger on legitimate tasks.
The broader read is simple. Anthropic is trading margin on cache reads to capture agent-heavy adoption and is meeting privacy demands with customer-controlled infrastructure. If the company executes on both, procurement friction drops and proof-of-value timelines compress. That would push more organizations to move live projects onto Claude Fable 5.1 sooner rather than later—and budget with greater confidence.
Related reading: Copilot • OpenAI • Productivity & AI