Anthropic shipped Claude Fable 5 this week, its first publicly available Mythos-class model, and the release is already reshaping the Claude app around it. Fable 5 posts a jump of more than 10% over Opus on several benchmarks, but it arrives with hard limits: prompts touching cybersecurity, biology, chemistry, and model distillation are blocked by safety classifiers. The model is bundled into paid Claude plans until June 22, after which it moves to extra usage billing while Anthropic works to fold it back into subscriptions.
Those restrictions appear to be driving the first of several unreleased features spotted in a recent iOS build. A new settings toggle lets Claude hand a conversation to a different model when Fable's guardrails fire, routing flagged prompts to Opus instead of surfacing an error. This mirrors the fallback behavior Anthropic describes on the API side, but as a user-facing control, it would make the classifier system far less disruptive for subscribers who hit those boundaries in ordinary work.
Voice mode looks set to gain a model selector as well. Anthropic is preparing a multi-language voice rollout, and the build now references switching between Opus, Sonnet, Haiku, and even Fable as the model orchestrating the TTS pipeline, tool calls, and related tasks. Voice currently relies on Haiku 4.5, which hasn't been upgraded in some time, so the selector, alongside an expected Haiku 4.5, would address a real capability gap.
The third discovery ties Claude Code to Anthropic's Managed Agents platform. Cloud-managed agents configured on the platform would become selectable in Claude Code, allowing users to prompt those environments from the app and delegate work to purpose-built agents. The wiring is not yet functional, but orchestrating a fleet of specialized cloud agents from a phone would extend the direction Anthropic set with Claude Code on the web and Cowork's task assignment. None of these features have been announced, and timelines remain unclear.