It's the first time in the history of the world.
Recently, Anthropic and OpenAI released a major upgrade to the voice mode simultaneously!
This is Claude.From just running Haiku upgrades to the Opus 4.8 and Sonnet 5 series, you can not only tweak Gmail, calendar, Slack in conversation, but also expand to 11 languages, but no Chinese.
On the other side of OpenAI, GPT-Live is fully operational.Moving directly to the macOS and Windows desktops, moving your mouth, commanding an entire team of Agents at the same time, and supporting Chinese.
Upgraded by Claude Voice
Prior to this, Claude's voice pattern could only run on Haiku, the smallest model.
You talk to it by voice and it understands, but a little bit more complicated problems can be easily overturned.
Right now,Opus 4.8 and Sonnet 5It's all up.
You can switch models in the middle of the conversation. The system automatically calls the selected models.The fastest version.Note that this is the fastest version, not the full version of reasoning.
The good news is that voice patterns will change.By default, load the model you used in your last text chat.The text and the context between the words are continuous and uninterrupted.
In terms of interaction, there are two models:Hands-freeIt's a great way to get your attention.Push-to-talkIt's the end of the story, the end of the story.
Of course, the most important thing is "tool call".
In voice conversations, Claude can now connect directly with you.Gmail、Google Calendar、Slack、Google Docs、Canva。
In a nutshell, it works for you by flipping emails, checking calendars, reading documents, sending messages.
Robert Bye, product manager at Anthropic, gives a few examples:
• Help me summarize Slack what I missed in the last few hours.
• Send a summary of the new project PRD to Nick on Slack.
It's the end of the story, the end of life.
Technically speaking, the tools in speech are calls and text chats.The same Connectors architectureIt's not a castration, and every time I call you, I ask for your permission.
All packages are available, but free users can only use Haiku plus one connector.Loans in regular quotasToday, the entire platform announced its launch.
11 languages without ChineseIn terms of language, Claude's speech pattern is supported this time.11 languagesEnglish, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Brazilian Portuguese, Latin, Spanish, European Spanish.
I've seen it all over again, and no Chinese.
For this part of the user, you have to keep typing.
And that's what Claude is.It doesn't automatically detect what language you're speaking.。
You can activate the opening to cut it, such as "Please talk to me in Japanese", or go to the settings and manually select it.
Interestingly, a developer has previously extracted a language list from Claude App's code, containing 18 languages, in which it says explicitly:Chinese。
As a result, 18 varieties and 11 varieties of Chinese were cut directly.
ChatGPT Voice is dead on the desktop
In contrast, the voice of ChatGPT is:A whole new architecture from the ground up。
It was just released on July 8th.GPT-LiveIt's an all-industrial voice model.
The so-called double work is that the model can handle both input and output audio streams at the same time.It's about making a dozen judgments every second.You don't have to wait until you're done.。
So GPT-Live will listen to you in the middle of what you're saying, "Wow", "I get it", like a living person.
And it's based on the fact that there's aTwo-tiered architecture。
The front stage is a slightly low-latency voice model, responsible for real-time conversations, turn-management, interruption responses, all things that require a millisecond response.
The backdrop isGPT-5.5He is responsible for complex reasoning, web searches, Agent tasks.
When you ask a question that needs a lot of thought, GPT-Live assigns the task asynchronously to GPT-5.5 in the background, and then it keeps talking to you, waiting for the results to come out and report.
It's called OpenAI."Delegated Reasoning"。
Specifically, OpenAI deferred dialogue and reasoned in depth, and was completely unraveling. You don't have to wait for the big model to come up with something to talk about.
There are also three types of deep reasoning:InstantAs soon as possible.MediumThe secondary reasoning.HighIn-depth reasoning, which corresponds to the different grades of GPT-5.5 respectively.GPT-Live-1 mini。
In terms of running score, this delegation structure is really manageable.The voice mode is finally no longer a "reduced version chat".
GPQA (expert-level scientific reasoning) from the previous high-level speech model45.3% jumped to 84.2%。
BrowseComp ((Agent web search capability) from0.7% to 75.2%。
GPT-Live, which simulates the real multi-round telecoms customer experience scenario, is also fully compressed.
 |  |  |
With this set, OpenAI moved it directly to the desktop.
Starting today,ChatGPT Voice officially launched on macOS and Windows desktop.
You can set one.Global shortcutsIn any application, a button can trigger a ChatGPT Voice.
It's also available on macOSAppshots featureChatGPT can read directly the contents of your current active window as the context of the conversation.
For example, when Excel says "help me calculate Q3 comparisons", it can see your chart.
In addition, you can use voice commands in ChatGPT Work and Codex.Multiple Agents working at the same time。
One agent writes the code and the other agent searches the document.
GPT-Live speaks to you on the front end, and the back end starts, checks and switches multiple working threads at the same time, without having to go back to the keyboard.
Currently, Plus, Pro, Business, Edu, Enterprise packages are available.
You can feel the gap in 10 seconds.
Can you break it?
Claude's voice is rotational, and you're just going to think about it.
GPT-Live can be interrupted. If you've changed your mind by half, it stops immediately and follows your new ideas.
How can I use my heart more?
Claude only did one thing at a time and spoke with his voice.
ChatGPT Voice can talk to you at the end of the desktop, while controlling a computer in the background, commanding multiple agents.
"Help me run a PR while checking yesterday's CI diary", two things can start running at the same time.
Can you move?
Claude Voice can call Gmail through a connector, calendar these SaaS tools, but it doesn't do that.Can't touch your desk。
ChatGPT Voice is compatible with Computer Use on the desktop.Direct control of your computerIt can touch files, programs, terminals.
How deep can you think?
Claude's bottom line isOpus 4.8The voice mode is the fastest version of Opus, not the full-blooded logic version.
On the GPT-Live side, High files can be assigned asynchronously toGPT-5.5 ThinkingIt's the ability to reason deeply, but it's also the ability to summon near-blooded reasoning in a speech scene.
A time when there was no need for an input box
It's easy to see the difference between the two.
In turn, Claude uses your own Gmail, calendar, and documents to help you figure out what's going on.
GPT-Live, which works both ways, is mainly about talking to one person. It's disconnectable, it's multi-threaded, it's controllable, and it makes you forget it's AI.
For those who don't write code, Claude's door is much lower, there's e-mails to take care of, a schedule to line up, and the mouth is open.
OpenAI is a cooler and more extreme way to go, and you have to have a team of agents in your hands to do something for them.
It's like Greg Brockman said,"It would have been very unnatural to write like that compared to speech".
The plastic button that accompanied humans for 40 years is now officially pushed to the corner by both OpenAI and Anthropic.