Just now, Anthropic and OpenAI simultaneously released major voice mode upgrades!
On the Claude side, it has upgraded from only running Haiku to the full Opus 4.8 and Sonnet 5 series. It can now access tools like Gmail, Calendar, and Slack during conversations, and has expanded to support 11 languages, but Chinese is not included.
On the OpenAI side, the full-duplex GPT-Live has been directly deployed to macOS and Windows desktops. You can command a whole team of Agents just by speaking, and it supports Chinese.
Claude Voice Major Upgrade
Before this, Claude's voice mode could only run Haiku, the smallest model.
You could chat with it by voice, and it could understand, but it would easily struggle with slightly complex questions.
Now, Opus 4.8 and Sonnet 5 are both available.
You can switch models mid-conversation. The system automatically calls the fastest version of the selected model. Note, it's the fastest version, not the full reasoning version.
The good news is, voice mode will default to loading the model you last used in text chat, maintaining context seamlessly between text and voice, switching back and forth without interruption.
In terms of interaction, there are two modes: Hands-free continuous listening with automatic replies, and Push-to-talk where you hold to speak and release to end.
Of course, the most important feature is the "Tool Calling".
In voice conversations, Claude can now directly connect to your Gmail, Google Calendar, Slack, Google Docs, Canva.
Say a sentence, and it searches emails, checks your calendar, reads documents, sends messages for you.
Anthropic's product manager, Robert Bye, gave a few examples:
· Summarize what I missed on Slack in the past few hours.
· Send the summary of the new project PRD to Nick on Slack.
Say it out loud, and the task is done.
Technically, tool calling in voice uses the same Connectors architecture as text chat, not a watered-down version. It will also ask for your permission before each call.
All subscription plans can use it, but free users only get Haiku with 1 connector. The full Opus suite requires a paid plan. Usage is deducted from the regular quota. Public beta rollout across all platforms starts today.
11 Languages, No Chinese
Regarding languages, Claude voice mode now supports 11 languages: English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Brazilian Portuguese, Latin American Spanish, European Spanish.
Looked from beginning to end, no Chinese.
For these users, typing remains the way.
Furthermore, Claude does not automatically detect the language you are speaking.
You need to actively ask it to switch, for example, "Please speak to me in Japanese," or manually select it in the settings. If not switched, the default is English.
Interestingly, previously, developers found a language list in the Claude App code containing 18 languages, which explicitly included Chinese.
However, 18 became 11, and Chinese was directly cut.
ChatGPT Voice Invades the Desktop
In contrast, ChatGPT's voice has completely changed its underlying architecture.
It uses the GPT-Live model just released on July 8th, a full-duplex voice model.
Full-duplex means the model can simultaneously handle input and output audio streams, making dozens of judgments per second. It can respond without waiting for you to finish speaking.
So GPT-Live will interject with "uh-huh" or "I see" while you're talking, like a real person listening.
Looking deeper, it's based on a two-tier architecture.
The foreground is a lightweight, low-latency voice model responsible for real-time conversation, turn management, interruption response—anything requiring millisecond-level reactions.
The background is GPT-5.5, responsible for complex reasoning, web search, and Agent tasks.
When you ask a question requiring deep thought, GPT-Live asynchronously delegates the task to the background GPT-5.5, then continues talking to you, reporting back when the result is ready.
OpenAI calls this "Delegated Reasoning."
Specifically, OpenAI has completely decoupled conversation latency from reasoning depth. You don't need to wait for the large model to finish thinking before you can continue speaking.
Reasoning depth is adjustable in three levels: Instant (fastest), Medium (moderate reasoning), High (deep reasoning), corresponding to different tiers of GPT-5.5. Free users get GPT-Live-1 mini.
Judging from benchmarks, this delegation architecture really works. Voice mode is finally no longer a "downgraded chat."
GPQA (Expert-level scientific reasoning) jumped from 45.3% to 84.2% compared to the previous advanced voice mode.
BrowseComp (Agent web search capability) soared from 0.7% to 75.2%.
In the τ3-Voice Telecom benchmark simulating real multi-turn telecom customer service scenarios, GPT-Live also comprehensively outperforms.
Talk and Run Multiple Agents Simultaneously
With this foundation, OpenAI directly deployed it to the desktop.
Starting today, ChatGPT Voice officially lands on macOS and Windows desktop.
You can set a global hotkey to instantly invoke ChatGPT Voice from within any application.
On macOS, there's also an Appshots feature. ChatGPT can directly read the content of your currently active window to use as conversation context.
For example, when saying "help me calculate Q3 year-over-year" while looking at Excel, it can see your spreadsheet.
Furthermore, you can use voice to command multiple Agents in ChatGPT Work and Codex to work simultaneously.
Have one Agent write code while another looks up documentation, just by speaking.
GPT-Live talks to you in the foreground while simultaneously launching, monitoring, and switching multiple work threads in the background, no need for you to return to the keyboard.
Currently available for Plus, Pro, Business, Edu, and Enterprise plans.
The Difference You Can Feel in 10 Seconds
Can it be interrupted?
Claude voice is turn-based. It thinks only after you finish. Want to interrupt? Wait.
GPT-Live can be interrupted. Change your mind mid-sentence, and it stops immediately to follow your new thought. It remembers where you were before the interruption.
Can it multitask?
Claude does only one thing at a time—voice chat.
ChatGPT Voice on desktop can talk to you while simultaneously controlling your computer in the background, directing multiple Agents to do their respective tasks.
"Help me open a PR while checking yesterday's CI logs," both tasks can start running concurrently.
Can it act?
Claude voice can call SaaS tools like Gmail, Calendar via connectors, but it cannot touch your desktop.
ChatGPT Voice on desktop, combined with Computer Use, can directly control your computer. Files, programs, terminal—it can interact with them all.
How deep does it think?
Claude's trump card is Opus 4.8, still one of the strongest models for multi-step reasoning and knowledge synthesis tasks. However, the voice mode runs the fastest version of Opus, not the full-powered reasoning version.
On the GPT-Live side, the High setting can asynchronously delegate to GPT-5.5 Thinking for deep reasoning, actually allowing near-full reasoning capability to be called in voice scenarios.
The Era Where Even the Input Box Isn't Needed
Seeing this, the divergence between the two companies becomes clear.
The turn-based Claude focuses on using your own Gmail, calendar, documents to help you sort out your own stuff.
The full-duplex GPT-Live focuses on feeling like talking to a person. Interruptible, multi-threaded, computer-controlling, making you forget it's an AI.
For non-coders, Claude has a much lower barrier. Have emails to sort, schedules to arrange? Just open your mouth.
OpenAI's path is more exhilarating and geekier. You need to actually have a team of Agents and real tasks for them to do.
As Greg Brockman said, "Compared to voice, typing itself is particularly unnatural."
Those plastic keys that have accompanied humanity for forty years are now officially being pushed into a corner by both OpenAI and Anthropic.
References:
https://x.com/testingcatalog/status/2080401533728940372
https://x.com/OpenAI/status/2080378182469857576
https://x.com/testingcatalog/status/2080385285951521150
https://x.com/claudeai/status/2080376094939603366
https://claude.com/blog/think-through-hard-problems-in-voice-mode
This article is from the WeChat public account "New Zhiyuan," author: ASI Revelation, editor: Moses
