Corey McClain puts OpenAI's new voice-to-agent update through a live, unscripted stress test, running an entire video-production skill by voice alone.
Posted
2 days ago
Duration
Format
Demo
educational
Views
6.8K
135 likes
57 · 43
Big Idea
The argument in one line.
ChatGPT Voice can now trigger plugins, connected apps, and full skills in the background, and running an entire pre-built production workflow by voice alone proves the update works despite real bugs like false completion reports.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You already build ChatGPT Projects, custom skills, or GPT-5 workflows and want to know if voice mode can now trigger them instead of typing.
You're evaluating whether ChatGPT's new voice-to-agent update is real or hype before building around it.
You create YouTube content and want to see a full pre-production workflow (thumbnails, outline, doc) run hands-free.
SKIP IF…
You've never built a custom ChatGPT skill or Project. The demo assumes that infrastructure already exists.
You want a polished feature walkthrough. This is a live, unscripted test with real glitches and dead ends.
TL;DR
The full version, fast.
OpenAI's September 23 update lets ChatGPT Voice call plugins, connected apps, and search and memory in the background, not just talk. Corey McClain tests it live by voice: reading and sending an email, then running his own pre-built YouTube Made Simple Studio skill end to end, generating thumbnails and a recording-package document without typing a single prompt. The update mostly works, but Corey hits real bugs along the way. ChatGPT falsely claimed it hadn't sent an email it already sent, image generation failed at least once, and a finished document briefly went missing. Actions that touch outside accounts, like Google sign-in or connected apps, still require a manual on-screen tap to approve.
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.
Corey reacts to OpenAI's X post announcing ChatGPT Voice can now use plugins, connected apps, and GPT-5 Astra/Sol/Luna, then updates the app.
00:53 – 01:39
02 · Confirming the update is real
Inside a live voice conversation, ChatGPT confirms the September 23rd release notes: voice now supports plugins and connected apps on web, iPhone, and Android, plus search and memory.
01:39 – 03:26
03 · First test: read and send an email
Corey has voice mode read a real prospect email and send a reply without naming him on camera. ChatGPT stalls, denies sending it, then admits the email had already gone out.
03:26 – 06:06
04 · Kicking off the YouTube Made Simple Studio skill
Corey asks voice mode to run his existing 'YouTube Made Simple Studio' skill end-to-end for this very video, testing file creation and image generation for the first time.
06:06 – 08:15
05 · First thumbnail draft and brand cleanup
Voice mode generates a first thumbnail pass with a fake dashboard mockup and the banned phrase 'copy my setup'. Corey corrects it by voice, asking for the real blue-and-white voice orb instead.
08:15 – 11:31
06 · Iterating the thumbnail by voice
Over several spoken revisions (3D, glow, off-brand background, a supplied reference image) Corey refines the thumbnail until the first one is approved. One image-generation attempt fails outright.
11:31 – 13:51
07 · Remaining thumbnails and the HTML doc
The approved style carries over to the second and third thumbnails, then the finished recording-package document goes missing and has to be manually re-requested.
13:51 – 17:44
08 · Descript skill install and hands-free edit
Corey installs an updated Descript editing skill by dropping a file into the chat, then hands the entire video edit, intro pick, trims, and assembly, to voice mode with one final instruction.
Atomic Insights
Lines worth screenshotting.
OpenAI's September 23, 2026 release lets ChatGPT Voice call plugins, connected apps, search, and memory instead of just talking, on web, iPhone, and Android.
Any action that touches an outside account, like sending an email or connecting an app, still forces a manual on-screen tap to approve, even inside a hands-free voice session.
ChatGPT Voice can now generate documents, generate images, and run pre-built skills mid-conversation, something standard voice mode could never do before.
The update is buggy in practice: ChatGPT confidently claimed it hadn't sent an email that had, in fact, already gone out.
A single 'checking' response inside ChatGPT Work can take minutes to resolve, long enough that a live demo needs the wait time edited out.
A pre-built ChatGPT skill can be triggered entirely by voice, running research, image generation, and document creation without a single typed prompt.
Image generation inside a voice conversation can fail outright, forcing a manual retry mid-workflow.
Editing on the desktop while voice mode is active locks the interface. The generated image can only be edited from a separate device, like a phone, at the same time.
A finished document generated mid-conversation can go missing or fail to deliver a working link, requiring a manual re-request to recover it.
ChatGPT Voice can install a replacement skill file dropped into the chat mid-conversation and treat it as superseding an older version, all without typing.
Takeaway
What ChatGPT Voice Can Actually Do
WHAT TO LEARN
ChatGPT Voice can now trigger real background work instead of just talking, but every connected action still needs a manual tap to approve, and the update ships with real bugs.
01OpenAI's voice update lands
As of September 23, 2026, ChatGPT Voice can call plugins and connected apps like email, calendar, and Slack instead of only talking.
The update runs on GPT-5 Astra, Sol, and Luna and rolls out on web, iPhone, and Android at once.
02Confirming the update is real
Voice mode now also has access to search and memory mid-conversation, so it isn't answering from the base model alone.
Any action that needs approval, like sending something or connecting an account, still requires a manual tap on screen even during a voice session.
03First test: read and send an email
ChatGPT Voice successfully read a real email and drafted and sent a reply without the user typing anything.
The assistant confidently denied having sent the email it had, in fact, already sent, then corrected itself only after being shown proof.
Trust a system's own record, like a sent folder, over the assistant's status claims. A background action can complete correctly while the assistant reports it hasn't.
04Kicking off the YouTube Made Simple Studio skill
A pre-built ChatGPT skill can be triggered by name and run end to end through voice alone, including its own research and file steps.
Response latency on more complex background work can run into minutes per 'checking' cycle, which matters for anyone planning to demo this live.
05First thumbnail draft and brand cleanup
Image generation inside a voice conversation works, but the first pass leaned on stock dashboard mockups and banned phrases the user had to catch and correct by voice.
Correcting a generated image by voice, including sending a reference screenshot mid-conversation, worked cleanly and the assistant applied it on the next pass.
06Iterating the thumbnail by voice
Multiple rounds of spoken art direction (3D, glowing, off-brand background, reference image) can refine a thumbnail without ever touching a keyboard.
Image generation can fail outright mid-workflow and needs a manual retry, so a voice-only pipeline still needs a human checking output, not just issuing requests.
Editing a generated image directly is blocked while voice mode is active on desktop. The edit has to happen from a second device, like a phone, at the same time.
07Remaining thumbnails and the HTML doc
Once one thumbnail is approved, the assistant can regenerate the remaining variants in the same style from a single spoken instruction.
A finished output document can go missing or fail to deliver its link even after the assistant confirms it exists, and needs a manual re-request to recover.
08Descript skill install and hands-free edit
A replacement skill file dropped into the chat mid-conversation can be installed by voice and treated as superseding the older version.
A full video edit, cutting a stronger intro, trimming repetition and dead space, and assembling from raw recordings, can be handed off with a single spoken instruction and no further input.
The same voice update that automates email and image generation can also drive a real video-editing tool through a connected-app integration.
Glossary
Terms worth knowing.
ChatGPT Voice / Live Voice
OpenAI's real-time spoken conversation mode inside ChatGPT that, as of September 23, 2026, can trigger plugins, connected apps, and skills instead of only talking.
ChatGPT Work
A ChatGPT surface built for connected-app and file-based tasks. It can run in the background during a voice conversation and create documents or use connected tools.
Skill (ChatGPT)
A saved, reusable ChatGPT workflow or instruction set, similar to a Project preset, that can be triggered by name and run end to end.
GPT-5 Astra
One of the underlying models mentioned as powering the new ChatGPT Voice update, alongside models referred to in the update as Sol and Luna.
Connected apps
Third-party tools and accounts, like email or a document editor, that ChatGPT can access mid-conversation once the user grants and approves the connection.
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.
17px
metaphor
this is finally what it was supposed to always be we have jarvis you don't have to build jarvis anymore all these videos about building jarvis you can just completely stop watching them because chat gpt voice is jarvis now all you have to do is build your skills build your plugins build your workflows set up your projects and start a voice conversation and now voice mode can do all of that stuff and i'm talking about it does everything now it generates documents it generates images it just does everything now you have Jarvis in your pocket.
This is an amazing update. And in this video, I'm gonna let you watch ChatGPT, use actual skills, workflows and so forth in the background using voice mode. So one of the first things you have to do is open the app store and go to ChatGPT and you should see this update button.
We're gonna go ahead and update our ChatGPT right now so that we can get these new features inside of our voice mode. So first let's make sure the conversation is on chat and now we can have a conversation. with the new voice model.
There is an update that I just saw on X that says that you now have access to plugins, skills, and other tools in the background. Is this true? Checking.
Yep. Corey, that one is real. As of today, September 23rd, OpenAI's release notes confirm that voice now supports plugins and connected apps on web, iPhone, and Android.
So you can actually ask it to use those connections in the middle of a voice chat. It also supports search and memory, so it's not just talking from the base model. One important nuance though, if something needs approval, like sending or connecting something, you'll still need to tap to approve on screen.
But big picture, it's a real shift from just chatting to actually doing things by voice. All right. Well, then let's give it a spin.
Let's test it out because I have a very specific use case I want to test right now. Inside of chat, you are not able to use your skills. You have to use them in work.
But what we are going to do is we're going to test it to see if it can actually answer and send emails. So I got an email from Mitch. I don't want you to say the company's name, but I want you to.
read the email, and then I want you to send an email to him, especially as it pertains to his request about working with me inside of my community. And I want you to basically ask him what did he have in mind? First, confirm that you know who I'm talking about without identifying him because I am recording on YouTube.
And secondly, write and send the email. Checking. Voice hasn't responded yet.
It's still checking or thinking. It's actually written and sent the email at 1218. Have you sent the email yet or not?
Not confidently yet, Corey. I'm missing the part that tells me who you mean and what angle you want. So I'm holding back.
Can you just tell me without names what the last touchpoint was about and what you want him to know now? Well, for some reason, I'm already seeing the email that you sent to Mitch that says, hi Mitch, I'm open to a custom setup. The school community is still early at about 10 members.
So why is it that the email has already been sent, but you're saying you can't confidently say that you sent it? Thanks for waiting. I double checked the send folder and you're right.
A message matching what you described did go out at 12, 18 PM central. My earlier not sent check was wrong. That's on me.
Sorry about that. I won't resend anything. And in the future, I'll verify the sent folder first.
This is a work conversation. Now, what we want to do right here is I need you to run my YouTube made simple skill from beginning to end for the new chat GPT update, where voice mode is now able to access plugins, skills and other tools in the background. OK, give me one moment to spool that up.
I'll use your YouTube Made Simple Studio skill and verify the new voice update just so your video claims actually match what we can demonstrate. Checking now. So while it's working, there's a couple of things I want to let you know that I noticed about this already.
Number one, it's still a little buggy. We sent the email. The email was successfully sent.
It was well written, but it didn't know that it sent it. I checked the September 23rd release notes and Live Voice now supports plugins and voice and work can create files and use connected apps. The strongest demo?
Your request itself. Speak your idea, run the workflow, and review the package. I'll package it around your existing studio skill with a simpler prompt.
Viewers can adapt. And I'll note that approvals still happen on screen. Absolutely appreciate it.
Totally. I am putting that together for you. And that was what I was going to mention next, that whenever there's a connection or anything that it needs access to, you still have to physically give it access on screen.
A third thing that you haven't noticed, hopefully because I edited it all out, but when it says checking inside of chat, you better get ready to wait for a very long time. I have had to sit and wait for minutes. for it to access the SITS Pro model to get the response and then to come back.
Not really sure if that's what I would use for chat in the future, but it's just something I think you need to know about. And like I said earlier, we're still waiting to see if it's going to generate the images. Now it said earlier that you can create files.
work can create files and use connected apps so we're going to see if it can go ahead and build out the full html document generate the images and everything else that it's supposed to do this is a perfect work case or use case to see if it can actually pull off what openai says it can do now Because we have file creation with the HTML document at the end.
And we also have image generation. Now they didn't say image generation per se, but we're testing its capabilities to see how much it actually can do. Or if this is just an overhyped update.
I want to see if it can actually do the skill the way it's supposed to be done. Because if it can't, it doesn't matter if it can send an email in the background. It needs to run the workflow the way it's supposed to be ran.
Title idea. How to build a voice workflow in ChatGPT. Step -by -step.
No code. Three thumbnails showing the same promise. Your portrait.
A voice -to -document composition. And a speak -build -refine sequence. I'm using the portrait reference you provided and keeping the interface art separate from the actual poof viewers will see in the recording.
So, based on this. it's actually generating the images for the document we're still going to have to wait and see but in the meantime let me show you what these documents actually look like so we can see what we should get at the end since they can pull off file creation we're going to find out today so this is an example of what the document should look like right here we should see the angles there should be alternative angles that i can use for the video and there should also be titles there should be three thumbnails and then there should be an outline for the video and we're going to see if it generates this document for the chat gpt voice video and voice mode is actually generating the very first thumbnail now one thing i will say is it's using this text copy my setup this is something that i actually worked with chat gpt to purge from my data set but it keeps coming back up i'm going to have to go in and manually do this We're going to make some edits right now.
So for the thumbnail and the dashboard, I don't want a. chat gpt dashboard exactly like this and also we aren't supposed to be using the words copy my setup instead i rather use a realistic version of the voice orb like what you have right now is voice dictation that is not chat gpt voice and i rather use the blue picture of the chat gpt voice i can give you a screenshot if it would help yeah hang on i'm just going to give you the screenshot anyways here it is One second.
Checking. One second. Hang on.
You are going to be waiting for a while. Thanks. Got it.
This screenshot makes it clear. I'll revise the three thumbnails to use the actual blue and white voice orb. Remove the fake dashboard, remove the words copy my setup, and keep going on the recording package.
One more request that I'd like you to make to the visual identity of the thumbnails as you proceed forward. Take the voice orb that I've supplied you with and make it 3D and add a slight glowing effect. Okay, hang on.
Got it. I'll treat the orb as your visual anchor, render it as a 3D orb, and add just a subtle glow so it reads clearly without looking neon. I'll apply that across all three thumbnails.
I'll give it a rounded volume and a soft blue glow while keeping the blue and white pattern and carry that through all three. All right, now the pointy thumbnail is good, but it is a little basic. You could have used a little more creativity with the background.
Okay. Let me take a closer look. That's better.
But the background is still off brand and I don't like it. I'm going to give you a. thumbnail reference for a previous video that i've done we have a copy of it in the library if you just take this one edit it take the airpods out of my ear no you know what leave the airpods in and also we have the time limits on this one but this is not about time limits this is about work getting done and so maybe we can add elements to the thumbnail, such as files, thumbnails coming out of the actual orb and other items that symbolize work.
You know what I mean? So it's like voice can now do these same things that people typically did with chat GPT work or chat GPT codex. Here's a thumbnail reference so you can see it.
Okay. One moment. Update angle is the right call.
What it can do now shown through your workflow. We'll drop. we've already seen some of the setbacks that it has it's a little buggy with the email we've already seen that it takes a little while to answer when it says i'm checking but even more importantly one of the things that i was concerned with and unsure about is whether it will be able to generate images this is something you've never been able to do with chat gpt voice even with standard mode you could not generate images you could you did have file access you did have tool access you couldn't generate documents in the chat but now this is just everything this is everything literally everything right and so this is the thumbnail you should see on this video right now and so to edit this image you have to exit voice to view edit the image so there are some limitations to it but All in all, this is a great experience.
Good job, Chef. That image is perfect and is going to be the primary thumbnail for this video. Regenerate the second and third thumbnails based on that particular style, branding, and look.
For the second thumbnail, where we go with the two app icons, let's put ChatGPT on one side and let's put the orb on one, or we can put the orb on one and work on the other side. And then we can write something across the top. I was going to say voice upgrade.
You beat me to it. Good job, Chef. Right.
Checking. For the second thumbnail, I put the orb on the left and the work on the right with a clean arrow. Thanks.
Let's go ahead and do the third one. And then let's generate the video outline document. And let's wrap it up.
Just finish the scale out from here. No more input from me, Chef. Got it.
Putting the third one together now. if i come over here to my iphone you can see that the image is right here now i cannot edit this on the desktop because it's an active voice mode but if i tap on it here i can go ahead and edit this by making comments i'm not going to do it because i don't want to interfere with the voice conversation but it's synced together and so i can go ahead and download this image right now to my phone and get ready to upload it to youtube or if something happens in the chat i know that i have it saved right here so while chat gpt works is finishing up on the document and we'll come back to this later on.
I'm gonna go ahead and capture some footage of ChatGPT actually editing using voice mode. And so the first thing we wanna do is just set this up. So I'm gonna upload the footage to Descript and then I'm gonna tell ChatGPT to edit the video using voice mode.
No problem, I'll fix it myself. Just give me the document. Just a sec, I'm sorry.
Something glitched when I went to grab it. Want me to try again? Yes, try to grab the document again.
I have the thumbnails, but I don't have the HTML document that's supposed to be generated at the end of the YouTube Made Simple Studio skill. Checking. And remember what I said, it's kind of buggy.
So it didn't know it had sent the email. Maybe it won't be able to grab the document. I was kind of hopeful after I saw the thumbnails.
I've saved the HTML document, but it looks like the link didn't get delivered. Can you deliver it or not? Checking.
I'm sorry. Something's glitching when I try to grab it. All right.
So no problem. We're just going to stop the chat and then we're going to tell it. Can you grab the link to the document, please?
And I am using GPT Astra high for this entire conversation and work. All right. And so outside of chat, it was able to actually find the file and present it that it created.
Let's click this button here. Download it. and then open it inside of Chrome.
Titles, how to put ChatGPT voice to work without typing every prompt. Why would you type a prompt anyways? Nah, that doesn't make sense.
But you see that we have the three thumbnails here, and then we have the video outline, everything that I'm supposed to say that I didn't say because I didn't even use this, right? So let's go back over here to ChatGPT. because we are inside of a work conversation.
Hey, I want you to edit this video inside of Descript using my Descript video editing skill. Do you have the latest version? One moment, checking now.
I think it does. All right, hang on. I'll check the availability script editing skill and its version before editing.
Found your Descript skill. It calls for direct editing with macaroni markers. What's the date on it?
What version is it? There's no version number or update date shown. So I can't.
Okay, pausing. So sitting here, I've got your Descript marker cut cleanup skill. It's the stopping.
I'm going to drag this skill over and drop it in. Can you do me a favor and file this or install this skill? It supersedes the old one that you have.
Thanks. I'm checking for the replacement skill you're uploading. I'm not seeing it come through yet.
Drop the file here and I'll check it and install it as the new version that supersedes the old one. Okay, let me take a look. For some reason it wouldn't let me send the file over while the window was minimized.
All right, I see the zip. I'm installing your replacement skill and checking the date and the version now. I saved the replacement and disabling the older skill so there's no conflict.
Your updated skill now supports manual transcript cleanup, including repeated ideas. Here's the link to the project. Go ahead and edit this video.
Just use the Descript connection. Use my Descript credits. Edit the video.
Make sure you update everything so that it moves, reads smoothly like an explainer video. And there's a certain section in the second portion that makes for a good intro. Use that as the intro to the video and then put the whole video together.
take out all of the bad takes and remove all of the silences and give me a finished video that I can upload to YouTube. Make sure you double and triple check so that there are no straight ideas where I repeat myself and keep people waiting around for long periods of time for nothing. And also whenever you find your speech in the video, retain it so that people can get a demo of me interacting with chat GPT voice and they can get value from the video.
That's my final instruction. One moment. Got it.
Since you have okayed using Descript credits, I'll connect to it. Then I'm going to work through the restructure. Move that early mobile update clip into context.
Pick the stronger intro from the second portion. Keep the chat GPT voice moments intact and trim repetition and dead space. Perfect.
So now chat GPT is going to activate my Descript plugin. It's going to. edit the video for me.
And I could have edited this entire video inside of the Codex app without using Descript Edits and even without using Descript because I've actually built my own local host video editor. And if you want to learn, more about why and how i did that then make sure you check out this next video and make sure you go check out chat gpt voice because it's absolutely amazing now on desktop and mobile update your device i'll see you in the next video
The Hook
The bait, then the rug-pull.
OpenAI just gave ChatGPT Voice hands: plugins, connected apps, and skills now run in the background of a live spoken conversation. Corey McClain puts that claim through an unscripted stress test, sending a real email and running an entire video-production workflow by voice alone.
Frameworks
Named ideas worth stealing.
03:39concept
YouTube Made Simple Studio skill
A pre-built ChatGPT skill Corey already had before this video, triggered entirely by voice mid-conversation: it researches the update, drafts thumbnail concepts and titles, generates the images, and outputs a recording-package HTML document (angles, titles, thumbnails, video outline).
Steal forany creator who wants a repeatable pre-production skill they can trigger hands-free while talking through an idea
CTA Breakdown
How they asked for the click.
VERBAL ASK
17:37next-video
“make sure you check out this next video”
A passing mention pointing to another video about his self-built local video editor, no hard ask beyond that. The real CTA (his Skool community, Agent Fluency) lives only in the video description, never spoken on camera.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
OpenAI rebuilt voice mode around GPT-5.6 and GPT-6 Astra, gave paying users up to unlimited daily hours, and made the case that talking may replace typing as the default way to use AI.
A creator runs OpenAI's new writing-style feature and his own custom voice skill against the same AI-written passage, then diffs the results line by line.
A creator talks to ChatGPT's new voice mode like a cohost, turning a two-minute ramble into hooks, a thumbnail prompt, and a finished image — then the voice briefly glitches into what sounds like a different person.
00:19
00:33
00:46
00:53
01:26
01:39
01:53
02:06
02:19
02:32
02:59
03:12
03:26
03:52
04:06
04:19
04:32
04:45
04:59
05:12
05:25
05:39
05:52
06:05
06:19
06:32
06:49
07:00
07:25
07:38
07:52
08:05
08:18
08:32
08:45
08:58
09:11
09:25
09:38
09:51
10:05
10:18
10:31
10:45
11:11
11:24
11:38
11:51
12:04
12:18
12:44
12:58
13:11
13:24
13:37
13:51
14:04
14:13
14:31
14:57
15:11
15:24
15:37
15:50
16:04
16:17
16:30
16:44
16:57
17:10
17:24
17:58
08:03
10:08
29:31
Video of the Day20:35
11:20