Claude voice mode gets Opus and Sonnet

On July 23, Anthropic expanded Claude voice mode. The experience previously ran only on Haiku for speed; paid users can now choose Opus or Sonnet and switch models during a conversation. Claude also defaults to the model last used in text chat.

That change alters the role of voice AI. Haiku suited quick exchanges, while Opus and Sonnet can spend more effort on counterarguments, missing assumptions, and longer reasoning. Anthropic illustrates the shift with pitch practice, offer comparisons, and product roadmaps.

From talking to acting

The updated mode can reach tools a user has connected, including Gmail, Slack, Google Calendar, and Canva. A spoken review of a client pitch can become a one-page document; an email discussion can end with drafted replies.

Claude asks for permission before using a connected tool, which is an important boundary. It is not a complete safety system. Ambiguity is easier to miss in speech than in a written command, so calendar changes, email sending, and document sharing should still end with explicit confirmation.

More languages, but Turkish is not listed

The announced list includes English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Brazilian Portuguese, and two Spanish variants. Turkish is not on it. Claude also does not automatically detect the language; users must select it or ask to switch.

The beta is available on mobile, desktop, and web. Free users get Haiku, one connected tool, and the available languages. Paid plans can use expanded models and all connected tools. Voice conversations count against normal usage limits.

How I would use it

I would use this less for immediately assigning tasks and more for sharpening the task definition. Explain a product idea out loud, ask Claude to separate assumptions, risks, and unmeasured claims, then turn the result into text and review it before any action.

Voice accelerates thinking but can hide detail. A useful habit is to end with a written checklist of proposed actions. A stronger model makes voice more useful, but without careful confirmation design it can also make mistakes happen faster.