10:08Corey McClain · Tutorial
The ChatGPT Work Secret Nobody's Talking About
Regular ChatGPT and ChatGPT work run on two separate usage meters, not one -- and that split is the whole trick.
July 12thThe argument in one line.
ChatGPT Voice stopped being a lightweight fallback and became a full reasoning interface, routing hard questions to GPT-6 Astra mid-conversation while giving paying users up to unlimited daily hours.
Read if. Skip if.
READ IF YOU ARE…
- You're a ChatGPT Plus or Pro subscriber who wants to know exactly how many voice hours you now get and where the setting moved.
- You use ChatGPT for thinking out loud, planning, or brainstorming and want to know if voice mode is now good enough to replace typing.
- You're curious whether OpenAI's voice push signals a real hardware device, not just a software feature.
SKIP IF…
- You're looking for a technical breakdown of GPT-6 Astra's architecture rather than a usage walkthrough.
- You don't use ChatGPT's mobile or desktop voice features at all.
The full version, fast.
OpenAI restructured ChatGPT Voice around a single GPT-Live model that now silently calls GPT-5.6 or GPT-6 Astra when a question needs harder reasoning, and it raised daily voice limits (Plus: 3 hours, Pro $100: 15 hours, Pro $200: unlimited, all on a rolling 24-hour window). Voice is now three distinct modes, Live, Advanced, and Standard, controlled from Settings, not a dropdown, and users on a voice limit no longer get silently downgraded to a lightweight mini model. The video argues this makes voice viable for real work, walking, driving, brainstorming, coordinating a coding agent, then demos a real research conversation before speculating that OpenAI's next hardware device will be voice-first.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:27
01 · Cold open: the claim
The presenter opens with the thesis that ChatGPT Voice killed the keyboard, and teases a hardware speculation for later.
00:27 – 01:41
02 · New voice hour limits by plan
Breaks down the new daily voice limits: Go/Plus get 3 hours, Pro $100 gets 15 hours, Pro $200 gets unlimited, all on a rolling 24-hour window.
01:41 – 02:45
03 · Reading the demand signal
Shows a post from OpenAI's Tibo describing unprecedented demand for Astra, dismisses the 'scarcity trick' theory, then reframes: the hour count isn't the real story.
02:45 – 03:57
04 · The real upgrade: GPT-Live routes to GPT-5.6 and GPT-6 Astra
Explains GPT-Live's dual-channel, full-duplex audio, and that it can now silently call GPT-5.6 or GPT-6 Astra mid-conversation for harder reasoning.
03:57 – 04:48
05 · Message caps still apply
When GPT-Live escalates to GPT-6 Astra, that exchange counts against the same Pro message cap as text chat (200 on $200/mo, 50 on $100/mo); the old Instant/Medium/High intelligence selector is deprecated.
04:48 – 05:45
06 · Voice is now three products
Voice splits into Live (full-duplex), Advanced (needed for live video/screen share), and Standard (classic turn-based).
05:45 – 07:38
07 · How to actually switch voice modes
The old in-chat dropdown is gone. Switching modes now requires Settings, then Voice, then Model, where users can also set auto-start-with-voice ('Jarvis mode') and background conversations.
07:38 – 08:17
08 · No more silent downgrade to Mini
Plus and Pro users who hit a voice limit no longer drop to a lightweight GPT-Live Mini model, they drop to Advanced instead.
08:17 – 10:11
09 · The keyboard-killing thesis and use cases
The presenter's central argument: typing puts a 'toll booth' between thought and the model, so removing it via voice enables walking, driving, planning, brainstorming, decisions, and hands-busy work while directing AI, including inside Codex.
10:11 – 14:41
10 · Live demo: voice research on a book
A real GPT-Live conversation where the presenter asks ChatGPT to recall a half-remembered book, narrows it to 'The Art of Thinking' by Ernest Dimnet, gets a chapter outline read back, and finds a free copy, all hands-free.
14:41 – 15:32
11 · Hardware speculation: a voice-first device
Points to leaked images of a donut-shaped, hockey-puck-sized OpenAI device and floats the theory that OpenAI is building dedicated voice-first hardware, distinct from a phone.
15:32 – 17:58
12 · From GUI to CUI, and sign-off
Closes with the GUI-to-touchscreen-keyboard analogy (comparing it to the BlackBerry's physical keyboard becoming obsolete) to argue voice is the next default interface, then signs off.
Lines worth screenshotting.
- OpenAI raised ChatGPT Voice limits so far that a $200/month Pro subscriber gets literally unlimited GPT-Live hours on a rolling 24-hour basis.
- GPT-Live is not a separate cheap voice model anymore, it's a router that silently calls GPT-5.6 or GPT-6 Astra mid-conversation whenever a question needs harder reasoning.
- When GPT-Live escalates to GPT-6 Astra, that exchange counts against the same message cap as the Pro text plan (200 messages on the $200 tier, 50 on the $100 tier), so heavy voice use can burn Pro-model quota fast.
- The old three-tier voice intelligence selector, Instant, Medium, High, is deprecated. Reasoning effort now uses the same controls as text chat.
- Voice mode has been split into three distinct products: Live (full-duplex, natural conversation), Advanced (needed for live video and screen sharing), and Standard (classic turn-based voice).
- The dropdown that used to switch voice modes from the chat screen is gone. The only way to switch is Settings, then Voice, then Model.
- Plus and Pro users no longer get silently downgraded to a lightweight GPT-Live Mini model after hitting a voice limit, they drop to Advanced instead.
- GPT-Live has full-duplex, dual-channel audio, meaning it listens while it speaks, so it can detect interruptions and ignore background noise in real time.
- Voice mode now works inside the ChatGPT desktop app and Codex, so a coding agent can be directed and coordinated by talking instead of typing prompts.
- The presenter frames this as a GUI-to-CUI (conversational user interface) shift, comparable to how the touchscreen keyboard made the physical BlackBerry keyboard obsolete.
Voice mode is now a reasoning tool, not just a convenience.
WHAT TO LEARN
ChatGPT's voice mode quietly became a full router to the company's strongest models, which changes what's worth doing hands-free versus at a keyboard.
01Cold open: the claim
- A strong claim stated in one sentence at the very top gives a viewer a reason to keep watching before any evidence is shown.
02New voice hour limits by plan
- Usage limits on a subscription can be generous in raw hours while still being constrained underneath by a separate, stricter message cap on the model actually doing the reasoning.
- A rolling 24-hour limit behaves differently from a fixed daily reset, so read the fine print on 'daily' limits before assuming they work like a calendar day.
03Reading the demand signal
- When a company reads its own scaling pressure publicly, treat the stated reason (compute demand) and the cynical read (artificial scarcity) as separate hypotheses and weigh the evidence, don't default to either.
04The real upgrade: GPT-Live routes to GPT-5.6 and GPT-6 Astra
- A single voice model can now silently escalate a spoken question to a far more powerful reasoning model mid-conversation, so 'just voice mode' no longer means 'the weaker option.'
05Message caps still apply
- When a feature consolidates old manual settings into an automatic router, check whether that quietly removes a control you were relying on, not just complexity.
06Voice is now three products
- Splitting one feature into named tiers clarifies capability but adds a decision the user didn't used to have to make.
07How to actually switch voice modes
- A feature that used to live in a one-tap dropdown moving into a settings menu is a real usability cost, worth flagging even when the underlying feature improved.
08No more silent downgrade to Mini
- Removing a default downgrade path (no more silent drop to a weaker 'mini' model) is a meaningful reliability improvement that's easy to undersell next to flashier headline numbers.
09The keyboard-killing thesis and use cases
- The friction of typing is not neutral, it measurably changes which tasks people are willing to bring to a tool; removing that friction expands the tool's use cases rather than just making existing ones faster.
10Live demo: voice research on a book
- A live, unedited demo of a tool doing real work (recalling a half-remembered book, cross-checking a claim, finding a free source) is more convincing evidence of capability than a feature list.
11Hardware speculation: a voice-first device
- When evaluating a company's hardware rumors, converging signals (leaked device photos, public hints, a pattern of prior public speculation) are worth tracking even without formal confirmation.
12From GUI to CUI, and sign-off
- Interface shifts (keyboard to touchscreen, screen to voice) tend to be framed as impossible right up until they aren't, so dismissing a coming shift purely because current tech feels sufficient is a weak argument.
Terms worth knowing.
- GPT-Live
- OpenAI's current full-duplex voice model for ChatGPT. It listens and speaks simultaneously and can hand off harder questions to a more powerful text model mid-conversation.
- GPT-6 Astra
- A frontier-intelligence OpenAI model that GPT-Live calls when a spoken question needs deep, multimodal reasoning beyond what the voice model can handle alone.
- GPT-5.6
- An OpenAI reasoning model GPT-Live can call for search and harder logic, positioned between the base voice model and GPT-6 Astra in capability.
- Advanced Voice Mode
- The ChatGPT voice mode required for live video and screen sharing, distinct from GPT-Live's audio-only full-duplex conversation.
- Standard Voice Mode
- The original turn-by-turn ChatGPT voice mode, where the user speaks, waits, then listens, without full-duplex interruption handling.
- Rolling 24-hour limit
- A usage cap that resets continuously based on the last 24 hours of activity, rather than resetting once at a fixed clock time each day.
- CUI (conversational user interface)
- An interface where the primary input is natural spoken or typed conversation rather than buttons, menus, or a keyboard, positioned as the successor to the GUI.
- Codex
- OpenAI's coding agent product, which the video shows can now be directed and coordinated using voice instead of typed prompts.
Things they pointed at.
01:41linkTibo (@thsottiaux) post on Astra demand
10:46bookThe Art of Thinking by Ernest Dimnet
Lines you could clip.
00:00
“ChatGPT Voice just killed the keyboard.”
08:48
“It's like typing creates this small little toll booth between my thought and the model. But voice mode removes that toll booth.”
06:44
“This is going to be for the people who really want to be Tony Stark and get into Jarvis mode.”
15:33
“The digital keyboard or the GUI did away with the BlackBerry, because now there was a graphical user interface you could use.”
21:21
“If digital keyboards can replace physical keyboards in everyday use, is it so far-fetched that these models can become so good at understanding our intent that speaking to them can become the primary method of communicating?”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
ChatGPT Voice just killed the keyboard and before this video is over, I'm going to show you why I think they might be coming for the smartphone too. But before we get into all of that, look at what they just did to ChatGPT Voice. Oh, and they also moved something inside of Voice and I couldn't find it anymore, but I found it and it's going to save you a lot of time because you're probably going to be looking for it.
And this is exactly why I feel so comfortable saying that OpenAI killed the keyboard because ChatGPT Go gets three hours of GPT Live One Mini. They no longer have access to GPT live though. And plus users get three hours of GPT live one.
I believe it was one hour before and the $100 users get 15 hours of GPT live one and $200 users get unlimited GPT live one. Now, 15 hours is technically a limit. But if you think about the fact that you have to sleep, you have to eat, you have to go to work, maybe you have to talk to your wife, talk to your kids and do all other types of things, maybe even watch something on Netflix.
There is no way you're going to talk to ChatGPT for 15 hours straight, but it's available, but it's there. It's practically unlimited for the regular person. So at the high end of this, OpenAI is essentially saying, if you're a pro user, we want you to be able to talk to ChatGPT all day if that's what you want to do, if that's how you choose to get work done, or if that's how you choose to entertain yourself with it.
And these limits are on a rolling 24 hour basis. So it resets every single day. You don't have to worry about not having access to GPT Live 1 if you're a pro user.
That is big news. And the crazy part about all of this is right now, demand for Astra is so unprecedented that they're pulling all the levers possible to sustain the demand, but they've not seen anything like it now. And we went through very steep growth before.
Priority will always be to keep excellent service for existing users, but we might have to pause new pro subscriptions for a bit if this continues. People are signing up like crazy to use. Chat GPT.
Now, of course, you're going to have people who say it's a scarcity trick. You're always going to have naysayers and cynics. That's to be expected.
And as plausible as that may sound, OpenAI has been the most generous AI company from day one. So I don't see any basis for me even entertaining an allegation like that. I think they're genuinely just dealing with a lot of demand right now.
So compute is clearly a part of the story and it's something that's important to them. But what I care about is what this allows us to do. Because 15 hours, unlimited hours, that's not the real story.
GPT Live is a new voice model that is superior to their older voice models. And what makes GPT Live so special is its dual channel communication. It's speaking and listening at the same time at all times, whether you're talking or it's talking.
So it can tell when you interrupt it. It can also pick up on background noises and ignore them. And if you tell it to stop talking, it will stop talking.
It's a very good model. And I've had quite a few conversations with it when it caught me off guard. Very good model.
And the second biggest thing about the release of GPT Live is that it had the ability to take your message, go back to a higher intelligence, get the response, and then bring it back to you through the voice mode. And apparently it was using GPT 5 .5 to do this because now. GPT Live has access to GPT 5 .6 Soul and it has access to GPT 6 Astra, something I want to make very clear right now so that you don't get confused.
When you're using GPT Live, and let's just say you're on the pro plan, you have 15 hours or you have unlimited hours to use this voice model. Whenever it uses GPT -6 or essentially the Pro model or GPT -6 Astra, that counts against the number of messages that you're able to send to the model. Because as you know, the $200 plan can only send 200 messages to GPT -6 Pro and the $100 plan can only send 50 messages to GPT -6 Pro, if my memory serves me right.
So if you ask 50 or 200 complicated things within a day, and it has to constantly go to that model you can burn through your credit so the old way of changing intelligence levels from high medium and low is deprecated and it's fading out if you still see it there right now then you need to update your chat gpt app if you can't do it from the app store then delete the app and reinstall it and the new voice mode should pop up for you And so the old voice specific instant medium and high intelligence are being deprecated.
Now the model and reasoning controls line up with regular text chat GPT. When you start a conversation, you can easily hit the effort level and then choose the effort or level of intelligence you want to use for the conversation. And this gets confusing because voice itself also has several different modes.
So first, of course, we have GPT Live, the natural full duplex conversation that can now take your complex queries back to GPT 5 .6 Sol High or GPT 6 Astra and bring you back a response. But we also have the older advanced mode. And with advanced mode, you can use this for live video and screen sharing.
GPT Live does not have live video or screen sharing yet. It'll get there eventually, but it doesn't have it yet. And then of course there's old faithful, the standard voice mode, the classic turn by turn voice.
So in case you're in one of those situations where you want to say something and then just listen to chat GPT, that's going to be the model for you. But this is what tripped me up today. There used to be a button at the top of the screen where you could hit the down button and you could select whether you wanted to use live advanced or standard.
Well, that's not there anymore. So if you want to switch voice modes and you want to switch from GPT live to advanced, so you could do screen sharing or switch to standard, then you have to go to settings and then you have to go to voice. And once you get there, you can tap the model and you can choose one of these different models.
You can choose the language. You can also choose so that every conversation you have automatically starts with voice. That means that every time you open chat GPT and you tap on new chat, it opens.
up into voice mode automatically. This is going to be for the people who really want to be Tony Stark and get into Jarvis mode. And then you can turn background conversations on or off.
So if you're going to be doing something where you want live video or you want ChatGPT to be able to look at your screen and watch you do something on camera or something on your screen, then you're going to want to turn on advanced mode. Another way that you can turn on advanced mode is when you hit the plus button and you see the screen share or the live camera and you tap on those, ChatGPT will tell you that in order for you to use this, you have to use an older model and you choose continue and it'll switch you over to advanced mode that way as well.
And in case you're wondering, background conversations just means that you can keep talking to ChatGPT while you're using other apps. We've already talked about the usage enough, but I want you to pay attention to this. Plus and Pro no longer switch to GPT Live Mini after reaching a voice limit.
And there are a couple of things that this means. Number one, you're not going to hit a voice limit if you have unlimited hours. You're not going to hit it.
If you're a pro user with 15 hours, you are most likely not going to hit a voice limit. But let's just say you wake up, you put your phone on voice mode and they don't have a kill switch after the voice stays quiet for so many minutes and you manage to get to those 15 hours with voice activated. They're going to drop you down to another model.
but it's not going to be GPT Live Mini. It'll be GPT Advanced. The reason this kills the keyboard is not so much because I can sit in my bedroom or sit here at my desk and talk to ChatGPT all day, but I can go for a walk.
I can drive. I can plan. I can brainstorm.
I can make decisions. I can work on a car. I can go to the gym.
I can play basketball. I can do whatever I want to do. And if it's my choice, I have the option of talking to ChatGPT and working.
It's like typing creates this small little . But voice mode removes that toll booth. It moves that friction out of the way so I can think faster, form my ideas faster, go deeper and get more done.
And when you think about voice mode, please don't just think about your smartphone. This also applies to the ChatGPT Codex desktop app. You can literally use voice mode inside of Codex to orchestrate so many different actions and plans that you can work on at the same time from one place.
Just with your voice, you already have Jarvis. And I'm not saying that keyboards are going to disappear tomorrow. Because there are still some things like code, precise language, certain kinds of prompts and other items where the model and yourself benefit from having clear and cleaned up precise language.
But when it comes to driving things, building things, moving fast, engineering and building, voice mode is the way to go. And if you think about the way we get work done. In most environments, if you're a manager, you're not sending instructions down to your employees through a memo.
Sometimes, yes, but for the most part, you're going to verbally communicate it and then they're going to take action based on what you said. And this is how the models need to work. People need to speak to them using their everyday language and then the models turn around and take precise, accurate action.
All right, let's start a voice chat and see what it's all about. So what's on your mind tonight? What's something that's been tugging at you?
There's this book that I've read years ago called The Art of Thinking, and I can't get away from it because I really feel like there are some nuggets in that book that would shape the way researchers approach large language model training. And that would make the models smarter because it's all about how humans think, how the human mind works.
I remember reading it about 20, almost 30 years ago, and we talked about this before, but I can't remember the. two books that we narrowed it down to. I was wondering if you could remember those two books that we narrowed it down to, and then give me a summary of the difference between those two books and which one is most likely to have a meaningful contribution to the large language model community.
Let me check. Okay. So here's the thing.
Demnitz is about personal intellectual self -cultivation. Wallace is about a process model of how thought works with mechanisms and sequences. For LLM research, Wallace is more directly useful, not as a how -to, but as a hypothesis.
That framework pushes you to test whether models need differentiated processes for association, restructuring, incubation, and verification. What's funny is that might actually be the thing that you've been trying to remember all along. So tell me, does that name Wallace ring any bells?
Yes, it does. Can you tell me what country he's from? Is he from France?
I know he's not from France. He's from England. What about the other author, Demnit?
Checking. Demnit is French. What is the name of Demnit's book?
The Art of Thinking. And what is the name of Wallace's book? The Art of Thought.
It was Demnit's book. Now, give me an outline of Demnit's book. Sure.
Okay. Yes. And now that I'm looking at it, it makes sense why this book left an impression on you.
It's in four big parts across 15 chapters. First, he asks what thinking actually is, how to judge the quality of your own thinking, and whether an art of thought is even possible. Then he goes into what Bloch thought.
Annotation, social pressure, and just the noise of daily life. Then he shifts to supports for thinking. Solitude, reading with intention.
And this is the part I think you were remembering. He talks about elaborating data in the mind, going back over what you know, reflecting, writing things down, preserving thoughts, and developing them into original ideas. He ends with creative thought.
where ideas come from, how to get at your own ideas, and how to express them. And here's the thing. That middle section particularly, the elaboration of data in the mind, is the closest to your LLM hunch.
It kind of raises that question you're circling. Are we feeding models tons of information without an equivalent process for reflection, preservation, and recombination? that chance for real thought, that's much closer to the kind of thing that you might be getting at.
What about the other author, Demnit? Where can I find a copy of that book? Or just look on Amazon and give me a link right here in the chat.
So every time we tell it to search, it's using 5 .6. Here it is. This is where the author is listed as Ernest Demnit.
Well, this is interesting because I can get this for free. So here's the book right here. Got it for free.
I can read it whenever I get ready. Thanks to ChatGPT helping me out. Well, listen, that's all I wanted for tonight.
Talk to you later. Good night. And it even turns itself off.
So you notice throughout that long conversation, whenever I ask ChatGPT something, it would actually go in search. It would say, let me chat. And it would take a while to respond.
I don't have a problem with that because we humans, we speak before we think. And ChatGPT is going to think before it speaks. But if I can talk and have conversations like that all day about things I'm thinking about and kick ideas around, that's going to help me be more creative.
And now I just want to drop a little quick conspiracy theory on you guys. And I'm not sure this is going to happen. I don't have any proof for this.
All you have to do is a quick Google search and you'll see a lot of images for what is supposed to be opening eyes new device because they've constantly talked about how we're in this. where AI is still tied to your phone. But recently people have been hypothesizing that this is going to be the new device that OpenAI creates, this donut shaped item.
And it was about a year ago where I did a video when they first started talking about creating their own devices and hardware, where I said they should create their own smartphone. But when you think about voice mode and you think about how the digital keyboard or the GUI did away with the BlackBerry, because now there was a graphical user interface that you could use and the things that we used before just weren't necessary anymore.
This was so much more streamlined, so much more easier. The same thing is going to happen when we transition from a GUI to a QE. a conversational user interface.
When I can talk and use my computer, which is another reason why I think open AI, getting away from Sora, but focusing so heavily on computer use was the absolute smartest move because as agents become more capable of using the computer and using it as fast as GPT -6 Astra is and being accurate, there will literally be no ends to the work that it can do.
Especially when you combine that with a conversational user interface, you have to start thinking about the bigger picture. Imagine being able to talk out loud, not even to your phone, just another device. And then these agents in the background just go out and do the things you requested them to do.
The aspect of technology and its development is that the tech becomes smaller and it disappears. It becomes frictionless in the way that we use it. When you think about the first cell phones, how big and bulky they were with their antenna to what we have now.
you can see the progression how things have advanced and to think that what we have now is somehow the end all be all i think you have to be a little naive not to see that it's possible for us to make a progression to a voice first interface if digital keyboards can replace physical keyboards in everyday use is it so far -fetched that these models can become so good at understanding our intent that speaking to them can become the primary method of communicating and getting anything done on the internet or work at all so my keyboard is still here it's not going anywhere it's not broken but i will tell you this i know for a fact that i will be using it less than i was using it today if you got value out of the video make sure you hit the like button subscribe to the channel and as always take care have a great day and i'll see you in the next video
The Hook
The bait, then the rug-pull.
The title promises unlimited voice, but the real claim is bigger: OpenAI quietly turned ChatGPT Voice into a full reasoning interface, one that can escalate a spoken question to its most powerful model without the user ever touching a keyboard.
Named ideas worth stealing.
00:27list
The Voice Tier Ladder
- Go / Plus: 3 hours/day (GPT-Live mini or GPT-Live-1)
- Pro $100: 15 hours/day (GPT-Live-1)
- Pro $200: Unlimited (GPT-Live-1)
The new daily voice-hour allowance by ChatGPT plan, all resetting on a rolling 24-hour basis.
Steal forexplaining any tiered usage-limit pricing model to an audience
05:44list
The Three Voice Modes
- Live: full-duplex natural conversation
- Advanced: required for live video and screen sharing
- Standard: classic turn-by-turn voice
OpenAI split what used to be one voice feature into three named products with different capabilities.
Steal fordocumenting feature tiers when a single product splits into named modes
How they asked for the click.
VERBAL ASK
17:50subscribe
“If you got value out of the video make sure you hit the like button, subscribe to the channel.”
Single soft ask at the very end, no mid-roll pitch; the description carries a separate paid-community link (Agent Fluency) not spoken on camera.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
