Anthropic unveils Claude Fable 5.1 and Claude Mythos 5.1

New Models and Their Positioning

Anthropic announced the release of two new AI models called Claude Fable 5.1 and Claude Mythos 5.1.

The company states, "We’re introducing Claude Fable 5.1 and Claude Mythos 5.1."

Both models share the same underlying architecture but differ in the level of safety controls applied.

Claude Fable 5.1 is offered as a generally available product, while Claude Mythos 5.1 is limited to trusted access programs.

Mythos 5.1’s safeguards are tuned specifically for cybersecurity and life‑science applications, according to Anthropic.

The announcement emphasizes that the models target the most demanding coding and knowledge‑work tasks.

Cost, Data Retention, and Safeguard Improvements

Anthropic claims the new Fable 5.1 will cost about 25 % less than its predecessor for typical token‑based workloads.

The reduction stems from lower pricing on cache reads, where the model reuses previously processed inputs.

For highly agentic workloads, the company estimates savings could reach roughly 45 %.

Customers will also benefit from a zero‑data‑retention option, which Anthropic calls Enterprise Frontier Safeguards (EFS).

EFS stores data on cloud infrastructure that is entirely controlled by the customer rather than Anthropic.

The rollout of EFS to enterprise clients will begin in phases later this fall.

Until EFS is fully deployed, eligible users can already run Fable 5.1 with a strict zero‑retention policy.

Safeguard enhancements aim to lower false‑positive rates, especially in cybersecurity contexts.

The new safeguards reportedly block 60 % fewer false positives than earlier versions.

Anthropic notes that the improved model can now discover software vulnerabilities but is barred from generating exploit code.

In the life‑science domain, Anthropic has partnered with the U.S. government to create an access program for Mythos 5.1’s advanced biology capabilities.

The company says enrollment for scientific users will open soon.

Performance Benchmarks and Early Use Cases

Benchmarks released by Anthropic show Claude Fable 5.1 outperforms the prior Fable 5 across coding and long‑running problem‑solving tasks.

When configured for Low or Medium effort, Fable 5.1 delivers results comparable to or better than Fable 5 while incurring lower cost.

The model defaults to High effort in Claude Code and to Medium effort in Claude Cowork and on Claude.ai.

In the Terminal‑Bench‑Science 0.1 benchmark, Fable 5.1 achieves higher accuracy scores at reduced expense relative to its predecessor.

The public leaderboard reproduces Claude Opus 5 at 30.0 % and Claude Fable 5 at 21.4 % on the same test, while Anthropic’s internal setup records 29.0 % and 24.7 % respectively, both within statistical noise.

Similar gains appear on the Terminal‑Bench 4.0 suite, where Mythos 5.1 and Fable 5.1 show comparable scores when effort levels are aligned.

Anthropic attributes the remaining performance gap between the two versions to earlier, less precise cyber safeguards that intervened on certain tasks.

The company expects that ongoing safeguard refinements will narrow this gap further.

On the “Humanity’s Last Exam” benchmark, Fable 5.1 with tool assistance reaches a pass rate of 56 % at a modest cost, surpassing the prior model’s 50 % rate.

These results suggest that the new models could accelerate agentic scientific research, terminal coding, and multidisciplinary reasoning.

Why This Matters: The lower price, zero‑data‑retention option, and stronger safeguards of Claude Fable 5.1 and Claude Mythos 5.1 make advanced coding and scientific AI work more affordable and secure for enterprise and research users.