Claude Fable 5.1 cuts prices by a quarter and promises zero retention for enterprises

September 1, 2026. Anthropic launches Claude Fable 5.1 and Claude Mythos 5.1, billed as “the world’s most advanced models for coding and knowledge work.” 25%. That is the estimated price cut on typical workloads, reaching 45% on highly agentic use. June 9, 2026. Their predecessor Fable 5 had shipped, only to be pulled offline three days later under US export controls. Why it matters: this time Anthropic is not playing the benchmark-score card alone — it is advancing on price, data retention and safeguards.

One model, two safeguard levels

Fable 5.1 and Mythos 5.1 are the same underlying model, separated by a single thing: the level of safeguards. Fable 5.1 is generally available, while Mythos 5.1 is only reachable through trusted-access programs, with protections designed specifically for cybersecurity and the life sciences.

The positioning is unusual. Where most vendors sell distinct models by size or use case, Anthropic is selling two faces of the same network, and accepts that the safeguard gap explains a difference in behavior. The specifications remain ambitious: a 1 million-token context window and up to 128,000 output tokens, according to the public Claude timeline. Default effort settings vary by product: High in Claude Code, Medium in Claude Cowork and on Claude.ai.

The fight shifts to price and retention

The most legible news is not a score, but an economic concession. Fable 5.1 costs “an estimated 25% less than Fable 5 for typical workloads,” with the reduction landing on cache reads. On highly agentic work the savings climb “up to approximately 45%.”

The pricing mechanics are worth unpacking, because they reveal the target. A cache read is an input the model has already processed and stored for reuse — exactly what repeats in a loop during agentic work, where the same long context is re-submitted at every step. Cutting the price of that operation subsidizes long-running agents rather than one-off requests. Anthropic is anchoring its model in production workloads, not in demos.

The second lever is data retention. The new Enterprise Frontier Safeguards (EFS) system promises “complete privacy, the same as a zero data retention policy,” by storing data in cloud infrastructure controlled by the customer, not Anthropic. Rollout happens in phases starting this fall; until then, eligible customers can use Fable 5.1 with effective zero retention.

Safeguards that discover without exploiting

The third axis is the safeguards themselves, and the line is drawn in the right place. Anthropic says it cut false positives in its protections: in cybersecurity, the new safeguards block “60% fewer false positives than before.” Concretely, Fable 5.1 can now discover software vulnerabilities — but not develop exploits for them.

That distinction has practical consequences for a security team. Until now, frontier models often refused to produce even an analysis of vulnerable code, which made them useless in defensive code review. By allowing discovery while blocking exploitation, Anthropic opens the door to assisted bug-hunting while keeping a stop on weapon production. The open question is where the line actually sits: the boundary between “demonstrating a flaw” and “providing an exploit” is blurrier in practice than in principle.

What zero retention actually changes

The EFS promise is also best read in reverse. Until now, “zero retention” almost always meant “we delete your data after processing” — a contractual assurance that still let data transit the vendor’s infrastructure. EFS inverts the logic: data stays in infrastructure the customer controls, and Anthropic has nothing left to store or delete.

The nuance is commercial as much as technical. For regulated sectors — healthcare, finance, defense — the question is no longer “can you erase our data?” but “can you avoid ever seeing it?” That is precisely the wedge Anthropic is trying to open against rivals whose compliance story still rests on the promise of deletion.

A successor that buries Fable 5

The context weighs on this launch. Fable 5, released June 9, 2026, was taken offline on June 12 under export controls, then restored on July 1. With Fable 5.1, Fable 5 becomes a “legacy” model, and the episode is closed — on the surface at least.

The published benchmarks tell the same story as the pricing concessions. On Terminal-Bench 4.0, CursorBench 3.2.0 and Humanity’s Last Exam, Fable 5.1 beats Fable 5, but the argument being pushed is the accuracy-to-cost ratio: at Low or Medium effort, Fable 5.1 achieves results “similar to or better than Fable 5” at a much lower cost. The frontier is no longer only about which model is best, but which model is best per dollar spent.

Transparency as a sales pitch

The same day, Anthropic published the system card for Fable 5.1 and Mythos 5.1: the transparency artifact documenting safety evaluations, limits and governed use cases. In biology, an access program developed in partnership with the US government governs the advanced capabilities of Mythos 5.1, with enrollment expected to open to scientists soon.

For a buyer, the practical consequence is immediate: Mythos 5.1 is not purchased, it is requested. Access runs through a trusted program with eligibility criteria, where Fable 5.1 is available to everyone. The segmentation echoes the classic security-products model, where the most sensitive capabilities are never handed to the general public.

Why the effort dial matters

Underneath the pricing sits a quieter change: Fable 5.1 ships with High effort by default in Claude Code, but Medium in Claude Cowork and on Claude.ai. Effort is the dial that trades thinking time against token spend, and the defaults now encode a bet about how each surface is used — coding agents get the full reasoning budget, while the chat and cowork surfaces are tuned for cost.

For a team running Claude Code at scale, that default has a real dollar value. If the same workload that once needed High can now hit “similar or better” results at Medium or Low, the savings compound with every agentic run. The benchmark-to-cost curves Anthropic published are, in effect, the argument for re-testing your effort setting rather than accepting the default blindly.

Taken together, the three moves — cheaper cache reads, customer-side data, and a discover-but-do-not-exploit line — point the same way. Anthropic is betting that the next buyers of frontier models will be the ones whose procurement depends on compliance, not just on leaderboard scores.

Verdict

If you are evaluating a model for coding or research agents, test Fable 5.1 on your real workloads: the 25% to 45% cut on agentic use changes the economics, and the High default in Claude Code should be weighed against the Medium default in Claude.ai for your use case.

If you face compliance requirements, the EFS promise — customer-side data, zero retention — is worth a demo the moment it is available. It is the most differentiating argument of this generation.

If you work in security research, watch the “discover without exploiting” line: it opens real defensive use, but its exact perimeter still needs to be tested on edge cases.

References

• Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1, September 1, 2026
• Anthropic — Claude Fable 5.1 and Claude Mythos 5.1 system card
• Anthropic Help Center — Release notes, September 2026
• Anthropic — Enterprise Frontier Safeguards
• ETTAYEB — Anthropic sets a hardware standard for its models (MHS)