Anthropic appears to be testing a meaningful upgrade for Claude Voice Mode: the ability to run spoken conversations with Opus or Sonnet instead of relying exclusively on Haiku. The change has surfaced in a feature-flagged interface rather than the public product, so it remains an unconfirmed preview rather than an announced release.
The apparent upgrade is more substantial than simply adding names to a settings menu. Earlier versions of the hidden model picker reportedly displayed multiple choices without changing the model behind the conversation. More recent tests indicate that selecting Opus or Sonnet can route the session through the chosen model, although that behavior has not been verified in a public Claude release.
If Anthropic ships the feature, Claude Voice Mode would become less of a fixed experience and more like the regular chat interface, where the underlying model can be matched to the task. That could make voice more practical for conversations requiring deeper reasoning, longer context, or more involved use of Claude’s tools.
The model picker appears to be becoming functional
The unreleased interface presents three familiar choices: Opus, Sonnet, and Haiku. A separate control for adjusting the thinking level appears beside the model selection, suggesting Anthropic is exploring more than a simple switch between fast and capable models.
The distinction matters because the picker had reportedly been visible in testing for several weeks without reliably changing the model used by a voice session. Choosing Opus or Sonnet could still result in the conversation running through Haiku, making the setting largely cosmetic. The latest observed behavior suggests that limitation may be changing.
Because the interface appears to remain behind a feature flag, it should not be treated as a feature available to Claude users. There is also no confirmed release date, plan eligibility, regional availability, or pricing information. Anthropic could alter the controls or abandon the implementation before any wider launch.
Still, a selector that affects the active model is a stronger sign of product development than an inactive interface element. It suggests Anthropic is evaluating whether its more capable general-purpose models can support the latency and interaction demands of a spoken conversation.
Why Opus and Sonnet could change the voice experience
Voice assistants have to balance intelligence with responsiveness. A faster model can keep a conversation moving naturally, but it may be less suitable for requests involving several constraints, detailed analysis, or a long chain of reasoning. A more capable model may produce stronger answers while introducing additional processing time or higher usage costs.
Offering Opus, Sonnet, and Haiku as separate choices would let users make that trade-off themselves. Haiku could remain useful for quick questions and lightweight exchanges, while Sonnet could serve as the general-purpose option. Opus could be reserved for conversations where reasoning quality matters more than speed. Those roles are plausible consequences of the model picker, not confirmed product rules.
The upgrade could be especially useful for people who treat voice as an alternative input method rather than a basic hands-free assistant. Tasks such as talking through a complex document, refining an idea over several exchanges, or directing a tool-heavy workflow may benefit from access to a stronger model. Whether those tasks work well in practice would depend on latency, usage limits, tool availability, and how much context the voice session retains.
Early testing of the hidden experience also suggests that familiar conversational behavior remains intact. The system appears able to pause when the user interrupts and to tolerate longer silences without immediately taking over. Those observations are encouraging, but performance in a limited test cannot establish how the feature would behave across devices, network conditions, languages, or a larger user base.
The underlying voice architecture may stay conventional
The model options do not necessarily mean Anthropic is building a new speech-native Claude model. The test appears consistent with a more conventional pipeline in which Claude processes the conversation as a general reasoning model and a separate text-to-speech layer produces the spoken response.
That approach would allow Anthropic to improve the intelligence of Voice Mode by changing the reasoning model without rebuilding the entire audio system. It could also make new Claude models easier to add as they become available. The trade-off is that a multi-stage pipeline may behave differently from a system designed to process and generate audio directly, particularly around response timing, emotional delivery, and overlapping speech.
The final architecture has not been confirmed, and the identity of any speech provider should not be assumed from the prerelease interface. Anthropic could also change the backend before launch while leaving the visible model selector largely unchanged.
For users, the architectural distinction is less important than the resulting experience. A voice conversation must respond quickly, handle interruptions, preserve context, and produce useful answers. Adding Opus or Sonnet would only be an upgrade if Anthropic can maintain those qualities while running a more demanding model.
A public rollout remains uncertain
The hidden selector leaves several practical questions unanswered. It is unclear whether every Claude plan would receive all three models, whether voice conversations would draw from existing model limits, or whether the thinking-level control would remain part of the finished interface. Anthropic has not provided a firm public timeline for the change.
Other details may also determine how useful the feature becomes. Claude could allow users to change models during a conversation, require a new session for each selection, or automatically fall back to a faster model under certain conditions. None of those behaviors has been confirmed.
The cautious conclusion is that Anthropic appears to be actively testing broader model support for Claude Voice Mode. The available signs point to a picker that may do more than decorate the interface, with Opus and Sonnet potentially handling live voice sessions instead of defaulting to Haiku.
That would give Claude’s voice experience a clearer connection to the model choices available elsewhere in the product. Until Anthropic announces the feature or enables it publicly, however, Opus and Sonnet support should be viewed as a promising experiment rather than a finished upgrade.