Anthropic Expands Claude Voice Mode to Opus

TestingCatalog independently reported that choosing Opus or Sonnet routes voice conversations to those models instead of silently falling back to Haiku 4.5. However, its July 23 article described the feature as hidden behind a feature flag and not yet publicly available, while characterizing a broader release as likely in the following days. Anthropic had not published a formal announcement in the material reviewed by TestingCatalog.

That difference leaves the exact rollout status unclear. The two reports agree on the central technical change: model selection in Claude Voice can now determine the underlying reasoning model.

TestingCatalog reported that the voice interface exposes Opus, Sonnet, and Haiku, along with a thinking-level control. It described the stack as conventional text-to-speech rather than a speech-native model, with Claude handling reasoning and conversation management; the publication reported that spoken output appears to use ElevenLabs.

For ML practitioners, routing voice turns to different foundation models can make quality, latency, and cost explicit design variables in voice-agent workflows. More generally, products that expose model selection in conversational interfaces give teams a way to reserve stronger models for complex tool use or longer reasoning, while retaining lower-latency models for simpler exchanges. The reported connector support is particularly relevant where spoken interactions need to invoke existing enterprise tools rather than remain isolated chat sessions.