Anthropic rewired Claude's voice mode on Thursday so it runs on the company's full model lineup — Opus, Sonnet and Haiku — instead of quietly downgrading every spoken conversation to the smallest model. The release note is titled "Think through hard problems in voice mode," which is a fairly direct admission of what voice could not previously do.
What was actually broken
Claude's model picker was cosmetic in voice. Whatever a user selected, spoken conversations ran on Haiku 4.5 — chosen originally to keep latency down. Voice now defaults to the last model used in text chat, auto-selects the fastest available version within that model, and can be switched mid-conversation.
Tools by voice
The more consequential half is connectors. Voice mode can now reach Gmail, Google Calendar, Slack, Canva and Notion, which means composing and saving an email by speaking. Anthropic is currently the only major assistant that lets a user do that from voice — OpenAI's voice mode still cannot drive tools.
Tiers and languages
Free accounts get Haiku only plus one connected tool. Paid accounts get all models and all tools. Languages — 11 locale options covering ten languages including Hindi, Indonesian, Japanese, Korean and two Spanish variants — are available on every plan, free included, though there is no auto-detection. The feature is in beta across mobile, desktop and web; Anthropic says it "works best from your phone."
Still turn-based
This is a brains-and-tools update, not a speech update. Claude's voice remains turn-based — listen, pause, respond — while OpenAI's GPT-Live is full-duplex, speaking and listening at once. Interruption handling has not improved, and Anthropic published no latency figures. Rivals still sound more natural.