Anthropic’s new model release is pitched as an answer to customers demanding cheaper, less obstructive AI and more control over their data. But the company is pairing that opening with continued restrictions on its more capable Mythos model and on sensitive cybersecurity work.

On Tuesday, Anthropic introduced Fable 5.1 and Mythos 5.1, updates to its advanced Claude lineup. Fable is broadly available through cloud platforms and Anthropic’s API, while Mythos remains reserved for registered partners working in cybersecurity or life-sciences research.¹

The commercial pitch is straightforward: lower bills for heavy users. Anthropic says Fable 5.1 is typically about 25% cheaper than its predecessor and can be up to 45% cheaper on complex agentic tasks, largely through reduced charges for cached data.² A post shared by Addy Osmani said the model keeps Fable 5 pricing while offering “75% cheaper API cache reads,” better endurance on long tasks and more natural writing.Addy Osmani (addyosmani)³

The company is also trying to answer privacy concerns. Its Enterprise Frontier Safeguards, due to begin rolling out in the fall, are designed to let customers retain data on their own cloud infrastructure rather than send it outward. Anthropic said it has “never trained on enterprise data without explicit permission, and never will.”¹ Yet the service will still monitor for misuse, with customers deciding how that monitoring operates.

Safeguards are being loosened selectively, not discarded. Anthropic says Fable 5.1 has more precise protections that should be less likely to block ordinary biology questions, and it can now help identify software vulnerabilities. But penetration testing, exploit generation and binary-based vulnerability scanning will still be redirected to Opus models; Mythos retains its earlier biology restrictions.²

That caution is reflected in Anthropic’s own assessment of Mythos. Its system card calls the model a “slight regression on overall misaligned behavior compared to Opus 5,” even as it reports improvements over earlier Mythos and Sonnet versions.¹ The result is a familiar AI-industry bargain: more capability and lower cost, delivered with guardrails that remain firmly under the vendor’s control.