ResearchAudio.io

Anthropic Shipped a Voice Update Without Touching the Voice

Opus and Sonnet replace Haiku. OpenAI went full-duplex instead.

Anthropic updated Claude's voice mode on Thursday and changed nothing about how Claude sounds. The audio layer stayed where it was. What changed is which model sits behind the microphone, and what that model is allowed to touch.

Fifteen days earlier, OpenAI made the opposite trade. It rebuilt the audio layer from scratch so ChatGPT could listen and speak on the same clock, and left the tool story for a later release.

Two labs, two weeks apart, one product surface, opposite bets. If you are building anything with a microphone in front of it, the split is worth reading carefully, because both companies were solving the same constraint and put the seam in different places.

The chain everyone started with

Nearly every voice assistant shipped since 2023 runs a version of the same three-model chain. A speech recognition model turns your audio into text. A language model reads that text and writes a reply. A speech synthesis model reads the reply out loud. OpenAI confirmed this is exactly how ChatGPT's previous Advanced Voice Mode worked.

The chain has two structural problems. Latency compounds, because every hop waits for the one before it to finish. And prosody dies at the first hop: pitch, pace, hesitation, and sarcasm all live in the waveform, and text does not carry any of it. By the time the language model sees your words, it has no idea how you said them.

Both labs looked at that chain this month. They cut it in different places.

OpenAI collapsed the chain

On July 8, OpenAI released GPT-Live-1 and GPT-Live-1 mini, and retired Advanced Voice Mode as the default. Both are full-duplex models: they process incoming and outgoing audio at the same time rather than taking turns.

Think of the old design as a walkie-talkie and the new one as a phone call. On a walkie-talkie, one side transmits while the other waits. On a phone call, both channels are open the whole time, which is why you can cut someone off, and why they can murmur agreement without stopping you. Technically, the model keeps an audio encoder and a decoder running concurrently on a shared state, so the response is being shaped while your speech is still arriving.

That unlocks three behaviors the chain could not produce. Interruptions land mid-sentence instead of at a turn boundary. The model can back-channel, acknowledging you while you are still talking. And it can stay quiet for long stretches, absorbing context until it is called on, which is what makes live translation possible at all.

Here is the catch, and it is the whole story. A model small enough to run on a conversational clock is not the model you want doing multi-step reasoning. OpenAI's answer is to route out: hard queries go to text models like GPT-5.5 for search, reasoning, and agentic work, while the voice conversation continues uninterrupted. The full-duplex model handles the talking. A bigger model handles the thinking.

Distribution is split by plan: accounts without a subscription get GPT-Live-1 mini by default, and Go, Plus, and Pro get the full GPT-Live-1. OpenAI says more than 150 million people use ChatGPT's voice and dictation features weekly. The release has visible gaps: GPT-Live-1 does not support video or screen sharing yet, and in OpenAI's own Hindi translation demo the assistant carried a heavy American accent with a stiff, bookish register.

Anthropic upgraded the middle box

Before Thursday, Claude's voice mode routed every query through Haiku, the smallest model in the lineup, specifically to hold latency down. That worked for "what is on my calendar" and fell apart on anything requiring depth. It is the same tradeoff OpenAI hit, resolved by shrinking the model instead of restructuring the pipeline.

The July 23 release swaps that out. Voice mode can run on Opus, Sonnet, or Haiku. It starts with whichever model you last used in text chat, moves to the latest generation of that family, and runs the fastest version of it. You can change models mid-conversation from the picker, without naming a specific generation.

The second half is tool access, and it is the part that changes what voice is for. Claude's documentation lists Gmail, Google Calendar, Google Docs, and Slack as usable inside a voice session, alongside web search. TechCrunch reports Canva and Notion as well. So "move my two o'clock to Friday" and "draft a reply to that thread" become things you can finish by talking, rather than things you ask about and then go do by hand.

What did not change: the voice model itself. Anthropic has not publicly described its audio stack, and the design remains turn-based. Claude listens, pauses to think, then answers. Anthropic told Engadget the release is focused on intelligence and tool access, with more voice work coming later this year. TechCrunch notes the practical consequence, which is that users should not expect improved interruption handling from this update.

The fine print: accounts without a subscription are limited to Haiku and a single connected tool. Voice conversations count against normal usage limits. Claude Fable is not available in voice mode. And voice mode does not run inside Claude Cowork or Claude Code, both of which keep dictation but not two-way speech, which means it cannot reach the projects and skills you have configured there.

Three things to take from the split

What to watch

Neither position is stable. Anthropic said explicitly that more voice work is coming later this year, which reads as an audio-layer rebuild in progress. OpenAI's product lead described voice as a candidate primary interface for long-running agentic work, which is not achievable without the tool layer Anthropic just shipped. Each lab has publicly named the half it is missing.

For three years, voice assistants have been graded on how human they sound. The next round gets graded on what they finish.