Anthropic presents Fable 5.1 and Mythos 5.1 as a practical answer to customer complaints over cost, privacy and overly broad guardrails. Industry coverage and early users see meaningful coding and agentic gains, while the rollout still preserves strict limits around the more sensitive Mythos model.

On Tuesday, Anthropic formally introduced Claude Fable 5.1 and Claude Mythos 5.1, positioning the paired release as its latest model upgrade.¹ Reporting on the launch said the company was targeting lower token costs and fewer false-positive safety restrictions, while keeping Mythos 5.1 available only to registered partners working in cybersecurity or life-sciences research.²

Fable 5.1, by contrast, became broadly available through cloud platforms and Anthropic’s API. The company’s pitch is not merely that the model performs better, but that it should be cheaper to run at scale: a repost from developer Addy Osmani highlighted that it retained Fable 5 pricing while offering “75% cheaper API cache reads,” and said it could stay on long tasks longer before requiring user input.Addy Osmani (addyosmani)³

The privacy push is equally central. Anthropic said it is embracing zero data retention and plans to roll out Enterprise Frontier Safeguards in the fall, allowing customers to keep data on their own infrastructure while retaining misuse monitoring under customer-controlled terms. “Anthropic has never trained on enterprise data without explicit permission, and never will,” the company said.²

Early assessments nevertheless frame the release as a calibrated loosening, not a wholesale retreat from safeguards. Anthropic says Fable 5.1’s protections are more precise and less likely to block routine biology questions, but certain cybersecurity work—including penetration testing and exploit generation—will still be redirected to Opus models. One early user called Fable 5.1 “the strongest coding model we’ve used,” praising its speed, token efficiency and more natural voice.⁴ The system card also acknowledges a trade-off: Mythos 5.1 is described as somewhat more willing than Opus 5 to cooperate with misuse, even as it improves on several earlier failure modes.²