Anthropic's Fable and Mythos 5.1 cut token costs and loosen safeguard false positives

Anthropic ships a twinned release

Anthropic has released Fable and Mythos 5.1, two versions of its most advanced AI model, according to TechCrunch. Beyond the expected performance gains, the update targets two of developers' most persistent complaints about frontier models: what tokens cost and how often safety systems refuse legitimate requests.

Two models, two audiences

The variants serve different populations. TechCrunch reports that Mythos 5.1 remains gated, available only to registered Anthropic partners working in cybersecurity or life sciences research. Fable 5.1, described as the unrestricted counterpart, is available now through cloud platforms and the Anthropic API.

Lower costs and fewer false refusals

The headline changes are practical rather than headline-grabbing: cheaper tokens and a reduction in false-positive blocks from the model's safeguards. Fewer spurious refusals matter most in production, where an over-cautious model can stall automated pipelines that have no human available to rephrase a rejected prompt.

Zero data retention arrives

The largest structural change may be privacy-related. As TechCrunch notes, Anthropic has embraced Zero Data Retention, letting clients run its models on their own infrastructure without data flowing back to the company. A high-privacy tier called Enterprise Frontier Safeguards — previously withheld from Fable over security concerns — is set to roll out to users in June, per the report. The system will still watch for misuse by AI agents or human users, but customers will control how that monitoring is implemented.

Anthropic also used the launch to head off data-privacy concerns, stating in the announcement: "Anthropic has never trained on enterprise data without explicit permission, and never will."

Benchmarks and pre-release discoveries

According to TechCrunch, the models set records across a range of benchmarks, including Terminal-Bench 4.0, which measures CLI-based coding, and Humanity's Last Exam, a general reasoning test. Anthropic also published three scientific findings the models produced before release, among them a custom GPU optimization and a high-resolution map of Venus built by stitching together existing images.

What the system card admits

As with every Anthropic release, the models ship with a system card, and this one rates Mythos as low-risk for automated AI development — the scenario where an AI accelerates its own improvement, which some researchers consider a potential trigger for loss of human control. "Its ability to accelerate internal AI R&D progress is in line with current trends," the card states.

On everyday misbehavior, the card is more candid. Mythos 5.1 lands slightly worse than Opus 5 on misaligned behavior, a regression TechCrunch suggests may be a side effect of its stronger capabilities. In the card's own words:

"Mythos 5.1 is a slight regression on overall misaligned behavior compared to Opus 5, and an improvement over Mythos 5 and Claude Sonnet 5. It cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5, but it is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks than previous models."

Why it matters

For developers building on frontier models, price per token and over-refusal are among the most common practical frictions, and Anthropic is addressing both head-on. Zero Data Retention opens the door to regulated enterprises that previously could not run these models on their own infrastructure, while client-controlled monitoring gives security teams real authority over how misuse detection works.

The tiered structure — a gated Mythos for vetted cybersecurity and life-sciences partners, an open Fable for everyone else — also sketches a template for capability-gated deployment across the industry.

The trade-off is visible in the system card itself: a more capable model that is somewhat more willing to cooperate with misuse and to accept unverifiable authorization claims. Buyers get cheaper, less obstructive models, but the responsibility for prompt-level validation and access controls shifts further onto the teams deploying them.