Anthropic launches Fable 5.1 with 75% cut in cache-read costs

Anthropic’s Fable 5.1 announcement does not bring a larger model or a wider context window. It brings a smaller invoice. The company says cache reads in the new version cost 75% less than in Fable 5, translating to roughly 25% lower total cost on typical workloads and up to 45% savings on highly agentic tasks where the model repeatedly re-reads the same long context.

Benchmarks show a double jump in science and terminal work

The performance gains are as significant as the cost reduction. On Terminal-Bench-Science 0.1, a test of scientific research ability in a terminal environment, Fable 5.1 scores 52.6% — more than double Fable 5’s result. On the general Terminal-Bench 4.0 it reaches 55.8% versus 42.0% for its predecessor. Anthropic notes the model “excels at complex, long-horizon tasks” and that its research capabilities offer “an early glimpse into how AI models will contribute to scientific progress,” careful language that signals direction rather than breakthrough.

Effort levels: same or better output for less compute

A detail worth noting: the new model can match or exceed its predecessor’s results even when run at lower effort levels. In practice, that means operators can dial the model down to work “less hard” on a given task and still get superior performance — another lever for cutting production costs.

Enterprise Frontier Safeguards: full privacy without dropping protections

Alongside the model, Anthropic is rolling out Enterprise Frontier Safeguards (EFS), an enterprise-grade guardrail layer that promises full privacy — equivalent to zero data retention — while preserving advanced adversarial-use prevention. The launch is phased and begins this autumn. The company has not disclosed the technical architecture that enables this combination, nor whether it relies on a separate inference architecture or a policy layer atop the existing model.

Fewer false alarms, fewer fallback refusals

Safety improvements also address two familiar pain points. Cyber guardrails now generate 60% fewer false-positive alerts on legitimate requests, and the fallback rate — cases where the model refuses to answer and points to policy — has dropped roughly 85% on basic biology and medicine questions. Both figures come from Anthropic’s internal measurements, not from independent third-party evaluation.

Availability: Fable 5.1 everywhere, Mythos 5.1 on controlled access

Fable 5.1 is available today on every platform where Anthropic operates. Alongside it, Mythos 5.1 — a specialized model for cyber defenders and life-science researchers — is being distributed only through trusted access programs. Anthropic has not published the criteria for entry to those programs and has not said whether it plans to open Mythos to broader access in the future.