Anthropic Brings Opus and Sonnet to Claude Voice Mode

Anthropic has opened Claude’s voice mode to its more capable models, letting spoken conversations run on the Opus and Sonnet tiers instead of the lightweight Haiku model that has powered the feature since it launched in 2025. The company confirmed the change on Thursday, July 23, 2026, framing it as a way to make talking to Claude useful for real work rather than quick lookups.

According to Anthropic’s voice mode documentation, the feature now starts with whatever model a user last picked in text chat and runs its fastest version by default, so a session inherits the intelligence of the model behind it. Users can also switch between Haiku, Sonnet, and Opus in the middle of a conversation through a model picker. In practice, that opens up work the Haiku-only setup handled poorly: longer reasoning, feedback on how you phrase a client pitch, or tool-heavy requests such as drafting an email and updating a calendar slot by voice.

Voice mode can also reach into connected apps, including Gmail, Google Calendar, Google Docs, and Slack, so a spoken request can pull context from a user’s own accounts or take an action like rescheduling a meeting. That app integration is the clearest point of separation from OpenAI, whose recently updated voice mode changed how ChatGPT talks but still cannot use outside tools to get work done.

A product decision, not a new voice model

For anyone tracking what frontier labs actually ship, the notable part of this release is what Anthropic did not build. There is no new speech-native voice model here. Anthropic told Engadget the update “is focused on intelligence and tool access,” and that it is “continuing to invest in voice” with “more to share later this year.” The underlying pipeline stays turn-based: Claude listens, pauses to think, then speaks, and it reportedly relies on an outside text-to-speech provider such as ElevenLabs for the spoken output.

Routing a reasoning model into voice is less an audio upgrade than a capability one. A spoken request can now be handled by a model that plans and works through a problem, where Haiku was tuned to return a fast answer over a considered one. The change to what voice can do comes entirely from the model behind it.

That is a different bet from the one OpenAI is making. OpenAI has poured resources into a bidirectional, speech-native system, GPT-Live, that processes what you say and generates a reply at the same time, closer to the rhythm of a phone call. Anthropic is instead routing its strongest reasoning models onto a conventional voice stack and accepting the conversational limits that come with it. Because the voice model itself is unchanged, users are unlikely to notice smoother interruption handling or a more fluid back-and-forth. What changes is how well Claude can reason about what it is asked, not how naturally it sounds while doing it.

The trade-off maps onto a question Claude users already face in text: the most capable model is not always the right one for the task. Choosing a heavier model buys depth at the cost of speed, and voice conversations are especially sensitive to latency. Putting the choice in the user’s hands, rather than defaulting everyone to the fastest option, is the practical substance of the update.

What’s available, and what’s missing

The upgraded voice mode is rolling out in beta to all users across Anthropic’s mobile, desktop, and web apps. Because it is a matter of routing existing models rather than new infrastructure, the rollout does not wait on Anthropic shipping fresh audio systems. Free accounts are held to the Haiku model and a single connected app, with the Sonnet and Opus tiers reserved for paying subscribers. Multilingual input, added earlier this year, is included, though Claude still cannot detect a language switch on its own; users have to name the language they are about to speak, either out loud or in the voice settings.

The picker exposes Opus, Sonnet, and Haiku, but not Claude Fable 5, Anthropic’s most capable widely available model, which stays out of voice for now. That omission fits the release’s logic. This is about matching voice to the reasoning models most people already use day to day, not about pushing the frontier of what a spoken assistant can do.

The result reads less like a voice breakthrough than a statement of strategy. Anthropic is wagering that the value of a spoken assistant comes from the model doing the thinking and the tools it can reach, not from a purpose-built audio architecture. Its promised follow-ups later in the year will show how long that position holds as OpenAI and Google keep pushing speech-native systems of their own.