Anthropic Fable 5.1 ships with zero-retention path for enterprises

Anthropic released Claude Fable 5.1 this morning, its new flagship for coding and knowledge work, alongside a restricted variant called Mythos 5.1 and an enterprise data architecture it says delivers privacy "the same as a zero data retention policy." Cache reads drop 75% to $0.25 per million tokens, and Anthropic estimates typical workloads cost about 25% less than Fable 5.

Key facts

• Fable 5.1 and Mythos 5.1 are the same model at different safeguard levels; Mythos 5.1 is gated to vetted US cybersecurity and life-science organizations (Anthropic)
• Cache reads now cost $0.25 per million tokens, down 75%; input stays $10/M and output $50/M (Anthropic)
• Enterprise Frontier Safeguards (EFS) store customer data in infrastructure controlled by the customer, which Anthropic calls "complete privacy (the same as a zero data retention policy)" — a provider claim (Anthropic)
• EFS rolls out in phases "beginning later this fall"; until then, eligible enterprise customers can use Fable 5.1 with zero data retention (Anthropic)
• Cybersecurity safeguards block 60% fewer false positives than Anthropic's previous system, per the company's own testing (Anthropic)

What happened

Anthropic shipped Claude Fable 5.1 on September 14, 2026, available today on the Claude API (claude-fable-5-1), Amazon Web Services, Google Cloud, and Microsoft Azure. The company frames it as "the world's most advanced model for coding and knowledge work" and posts benchmark leads over Fable 5, Opus 5, and GPT-5.6 on Terminal-Bench 4.0 (55.8%), Humanity's Last Exam with tools (65.0%), and GDPval-AA v2 — all Anthropic-run evaluations, not independent results.

Claude Mythos 5.1 is the same weights with safeguards tuned for sensitive domains. Access runs through Anthropic's trusted access programs: a cyberdefense track and a biology access program developed with the US government, currently limited to US organizations. Anthropic says Fable 5.1 "can now be used to discover software vulnerabilities — though not to develop exploits for them."

The pricing change targets context-heavy work. Cache reads — the tokens a model re-reads from already-processed context, which dominate agentic workloads — fall 75% to $0.25 per million tokens. Anthropic's own four-week August 2026 usage analysis puts the effective cut at about 25% for typical workloads and up to 45% for highly agentic ones. Base prices are unchanged at $10/M input and $50/M output.

Why it matters

The privacy move is the bigger story for builders choosing where sensitive workloads run. Anthropic's new Enterprise Frontier Safeguards keep prompts and outputs inside cloud infrastructure controlled entirely by the customer, with Anthropic saying the result is "complete privacy (the same as a zero data retention policy) while still being state-of-the-art at preventing adversarial use."

That wording is Anthropic's own, and the guarantee boundary matters. EFS is a storage-location architecture: Anthropic's claim is that retention-equivalent privacy follows from the customer holding the data, not from a policy promise. What would verify it is the published EFS terms, whether any logging exists for abuse review, and third-party audit — none of which exist yet, because EFS "will be made available to enterprise customers in phases, beginning later this fall." Until it ships, Anthropic says eligible customers can run Fable 5.1 under zero data retention, which is itself a policy statement rather than a verified property. As our explainer on what zero data retention actually means in AI APIs lays out, ZDR claims are unverifiable from outside without technical enforcement or audit.

The enterprise posture also now diverges sharply from Anthropic's consumer one. Days earlier, Anthropic updated its consumer terms so that Free, Pro, and Max users who opt in to model improvement can have conversations retained for up to five years; opted-out users stay on a 30-day window, and Commercial Terms customers remain no-training by default. Enterprise privacy is tightening toward ZDR-equivalence while consumer retention lengthens — the split procurement teams will be asking about, and the one our tracker on whether Anthropic trains on your data follows.

Background

Anthropic has competed on safety positioning since the Claude line launched, but data retention has been the gap enterprise buyers flag most. Zero-retention commitments have spread across the hosted-API market over the past year — OpenAI offers ZDR on qualifying endpoints, and OpenRouter added a provider.zdr request field this week — though every one of these remains a policy statement rather than an attested fact, as our zero-retention API roundup documents.

The Mythos 5.1 gating continues a pattern Anthropic set with tiered safeguard levels: capabilities it considers dual-use sit behind vetting rather than being refused outright. The biology access program, developed with the US government, is the clearest sign yet of frontier labs routing sensitive capability access through government-coordinated channels.

For cost, the cache-read cut is a direct shot at the economics of long-horizon agents, where re-reading context dominates the bill. Builders comparing hosted options against self-hosted open weights should note the cut applies wherever usage is billed by token — API, Claude Code, and enterprise plans — but the base rates are unchanged. Our migration guide for teams leaving Anthropic covers the workload math in the other direction.

What's next

Watch three dates. First, the EFS phased rollout "later this fall" — the published terms and any audit commitments will determine whether the ZDR-equivalence claim is checkable. Second, Mythos 5.1 access expansion: Anthropic says it is coordinating with the US government to open access to more domestic and international organizations. Third, Anthropic says it is "continuing our work to bring many of the improvements of Fable 5.1 to the rest of our model family," which implies Sonnet and Haiku refreshes follow. Independent benchmark runs, once third parties get API time, will test the posted leads.