The safety architecture is interesting. When Fable 5 encounters a high-risk query in cybersecurity or biology, it blocks its own response and hands off to Opus 4.8, which delivers a safer answer instead. Every Fable 5 deployment is quietly running a less-capable model underneath as a backstop. So the safety mechanism here is capability reduction, with a whole second model sitting behind the curtain.
Then there's the dual-track release. Alongside the public launch, Anthropic is shipping Claude Mythos 5, with some safeguards removed, to a small group of cyberdefenders through Project Glasswing, in collaboration with the US government. Same model, two access tiers based on who you are and what you're doing with it. Which formalizes what's been implicit at every major lab: the most capable version of a frontier model has a guest list.
The WWDC timing is conspicuous. Bloomberg has reported that iOS 27 will open Siri to rival AI assistants including Claude. Launching your strongest public model on day two of Apple's developer conference is not accidental.
Pricing lands at $10/$50 per million input/output tokens, less than half the Mythos Preview cost. The free trial through June 22 is a clean conversion funnel. Once teams experience an 80% SWE-Bench Pro model, stepping back to Opus 4.8's 69% will feel physically painful.
Sixty days from "too dangerous for public use" to "try it free for two weeks." The safety work and the launch playbook are both doing exactly what they're designed to do, and those designs have very little to do with each other.
Source: Anthropic ・ TechCrunch