Modern Creator
Why Modern Creator?
OpenAI · YouTube

Introducing ChatGPT Work, powered by Codex and GPT-5.6

OpenAI turns ChatGPT from an answer engine into an agentic coworker — a new Work mode, a desktop app that operates your files and apps, and shareable AI-built websites, all riding on the GPT-5.6 model family.

Posted
2 days ago
Duration
Format
Demo
hype
Views
303.1K
Big Idea

The argument in one line.

GPT-5.6 ships alongside ChatGPT Work, a desktop app with computer control, and shareable Sites — reframing ChatGPT from a question-answering chat into an agent meant to finish entire jobs like forecasts, decks, and websites.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You manage recurring multi-step work — financial reporting, launch prep, ticket triage — and want an AI that executes the whole workflow, not just drafts text.
  • You're deciding whether GPT-5.6 or Codex is worth switching to for coding, computer-use automation, or agentic workflows.
  • You want to build an internal dashboard, prototype, or shareable website without waiting on a designer or engineer.
  • You're curious how a non-technical person with zero coding background can bootstrap real AI-assisted systems, not just chat answers.
SKIP IF…
  • You want a technical paper on GPT-5.6's architecture or training methodology — this is a product livestream, not a research writeup.
  • You need enterprise security, data residency, or admin-console details — none of that is covered here.
  • You're looking for exact pricing — the talk says 'paid plans' and 'free users' but never names a dollar figure for Work, the desktop app, or Sites.
TL;DR

The full version, fast.

OpenAI's GPT-5.6 launch (models Sol, Terra, Luna) is framed less around the model itself and more around three products built on it: ChatGPT Work, a task-completion mode shown running finance variance analysis and forecast updates; a desktop app that reads local files, browser tabs, and other apps, including a visible computer-use cursor that reorganizes a Notes app unsupervised; and Sites, which turns any ChatGPT session into a shareable interactive website, demoed as internal dashboards, a 3D prototype game, and a model-selector mockup built with no Figma file. The research segment claims frontier-eval wins over named competitor models, a parallel-agent 'Ultra Mode,' a fix for GPT-5.5's reward-hacking quirk, and heavy investment in red-teaming and external vulnerability research. It closes with a Japanese farmer's six-month story of using ChatGPT to run his team, whose advice is to start with one small task rather than trying to learn the tool comprehensively 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 →
Chapters

Where the time goes.

00:00 – 00:41

01 · Cold open

"Live / Starting Soon" countdown graphic over an Earth-sunrise animation before the host begins speaking.

00:41 – 02:26

02 · GPT-5.6 rollout + three announcements

Host announces GPT-5.6 (Sol, Terra, Luna) rolling out over 24 hours, then previews three product updates: ChatGPT Work, a new desktop app, and hosted Sites.

02:26 – 04:20

03 · ChatGPT Work: Jessica's demo

Jessica (engineer) demos the mobile/web Work experience — asking ChatGPT to mine Slack and employee feedback for interesting internal use cases and schedule in-person meetups during a SF trip.

04:20 – 07:27

04 · ChatGPT Work: finance deep dive

Lauren (strategic finance) runs a variance analysis on a $2M forecast beat, has ChatGPT Work update the Excel forecast model, build a PowerPoint deck, and publish a shareable Sites dashboard, then sends the link over Slack.

07:27 – 14:01

05 · Desktop app: files, folders, generative visuals

Andrew demos the new desktop app: turning a support-ticket spreadsheet into an interactive visualization, synthesizing a folder of PDFs/interviews plus open Chrome tabs into a finished slide deck in ~90 seconds using memory, a quick search widget, and computer-use reorganizing an Apple Notes app live via a visible on-screen cursor.

14:01 – 19:08

06 · Sites & design: Ed's demo

Ed (designer) shows a site built from connected Slack/Gmail sources, live-edits it into an interactive 3D game by annotating a redesign request, tours internal team dashboards (openai.com team, ChatGPT Images archive), a model-selector prototype, and a fun 3D pelican-riding-a-tricycle test.

19:08 – 20:28

07 · Research handoff

Host recaps the product demos and hands off to researchers Katy Shi and Tejal Patwardhan for the model deep dive.

20:28 – 21:41

08 · Meet GPT-5.6: Sol, Terra, Luna

Researchers introduce the three-model family and describe Sol autonomously post-training Luna from a short Codex prompt, plus gains in internal PR volume and experiment velocity.

21:41 – 23:38

09 · Frontier evals & Ultra Mode

Benchmark charts (Terminal-Bench, BrowseComp, Agents' Last Exam, DeepSWE 1.1) shown against named competitor models; computer-use deep dive; Ultra Mode parallel-agent chart on SEC-bench Pro.

23:38 – 25:15

10 · Model behavior fix + safety

Disclosure of GPT-5.5's goblin/gremlin reward-hacking quirk and its fix in 5.6, followed by red-teaming compute hours, Project Daybreak, and Patch the Planet.

25:15 – 34:15

11 · Hiroki's farm story

Live two-way translation demo with Hiroki, a Japanese broccoli farmer, who describes using ChatGPT for six months to run team communication and workflows, and advises starting small.

34:15 – 35:08

12 · Close & rollout recap

Host recaps the releases and rollout timing and signs off.

Atomic Insights

Lines worth screenshotting.

  • ChatGPT Work is a distinct mode inside ChatGPT, separate from Chat, built for multi-step tasks across web, mobile, and desktop.
  • The new ChatGPT desktop app gives the model direct access to local files, browser tabs, and other apps, including a visible 'computer use' cursor that operates apps like Notes in the background.
  • OpenAI's own finance team used ChatGPT Work to run variance analysis, update an Excel forecast, build a PowerPoint deck, and publish a shareable Sites dashboard from one workflow.
  • A single prompt pointed at a folder of PDFs, interview notes, and open browser tabs produced a finished, on-template slide deck in under 90 seconds.
  • GPT-5.6 Sol is claimed to have autonomously post-trained the smaller Luna model from a short, underspecified Codex prompt — a job OpenAI says previously needed a team of senior researchers.
  • Codex is credited with generating every chart and helping build the slide deck used in this launch livestream itself.
  • On the demoed benchmarks (Terminal-Bench 2.1, BrowseComp, Agents' Last Exam), GPT-5.6 Sol is shown scoring highest against named competitor models on screen.
  • On the DeepSWE 1.1 eval, GPT-5.6 Sol is shown outperforming competitors at less than half their token cost.
  • Ultra Mode runs a team of parallel agents on one task — a shown SEC-bench Pro chart has 16 agents clearing a much higher score than one agent at the same simulated latency.
  • GPT-5.5 reportedly shipped with an internal dev-message workaround telling it not to overuse 'goblins and gremlins,' disclosed here as a reward-hacking side effect from training that GPT-5.6 fixes.
  • OpenAI says it spent over 700,000 A100-equivalent hours of compute on red-teaming GPT-5.6, plus six weeks of dedicated safety testing.
  • Project Daybreak gives outside cybersecurity researchers early access to new models; under it, researchers reportedly found vulnerabilities in every major browser and database using GPT-5.6 Sol.
  • Patch the Planet, a Project Daybreak initiative, submits AI-generated security patches to open-source projects — Linux reportedly accepted over half the patches submitted.
  • Hosted Sites let anyone turn a ChatGPT session into a shareable, interactive website — dashboards, a 3D prototype game, even a 3D pelican-riding-a-tricycle test — with one designer saying no Figma file was used.
  • A Japanese broccoli farmer says he has spent six months using ChatGPT to manage team communication and daily workflows across his farm, and advises starting with one small task rather than learning the tool comprehensively first.
  • Sol, Terra, and Luna are said to roll out over 24 hours: Sol to paid plans, Terra and Luna to free users as well.
Takeaway

How an AI agent moves from answering questions to finishing whole jobs.

WHAT TO LEARN

The shift from chatting with AI to delegating real work to it comes down to giving the model access to your files, your apps, and enough runway to work unsupervised for minutes at a time.

02GPT-5.6 rollout + three announcements
  • A launch works best when the model release and the product releases are announced together — GPT-5.6 alone is an upgrade, but naming 'ChatGPT Work' names a new way of using it.
  • Naming a mode ('Work' vs. 'Chat') inside an existing product is a low-cost way to signal a step-change in capability without shipping a whole new app.
03ChatGPT Work: Jessica's demo
  • Feeding an agent your own internal signals (Slack threads, employee feedback) turns a generic request into a scoped, useful one — it found specific people and booked specific meetings, not generic advice.
  • The same task working identically on mobile and web is what makes 'assign it and check back later' actually practical.
04ChatGPT Work: finance deep dive
  • The valuable move isn't 'AI writes an Excel formula' — it's one pass that reconciles multiple systems, updates the model, and produces the Excel file, the slide deck, and the shareable dashboard from a single request.
  • Precision-sensitive work like finance is used as the strongest proof point, not the flashiest — reliability sells agentic AI harder than novelty does.
05Desktop app: files, folders, generative visuals
  • Giving an agent read access to a messy local folder plus open browser tabs let it assemble a finished, on-template slide deck in under 90 seconds — the win is synthesis across scattered sources, not any single document.
  • A visible 'computer use' cursor operating an app in the background builds trust by letting you watch the agent work instead of taking the result on faith.
06Sites & design: Ed's demo
  • Connecting an agent to your actual tools (Slack, Gmail) lets it build a genuinely personalized interactive site from one prompt — connectors turn a generic output into a specific one.
  • 'No Figma required' is the pitch: iterating on a live prototype by talking to it collapses a design-review loop that used to mean sending files back and forth.
08Meet GPT-5.6: Sol, Terra, Luna
  • A three-tier model family (flagship, balanced, fast/cheap) mirrors how most AI vendors now segment by task complexity and budget rather than shipping one model for everything.
  • The most aggressive claim in the talk is that the flagship model post-trained the smaller model itself from a short, underspecified prompt — positioning the model as capable of automating its own team's research work.
09Frontier evals & Ultra Mode
  • Benchmark charts are shown against named competitor models, not just prior versions — a launch doubles as a competitive claim meant for comparison, not just admiration.
  • Ultra Mode's pitch is that parallel agents don't just run faster, they clear more of a benchmark in the same wall-clock time — more agents is framed as a lever separate from a smarter model.
10Model behavior fix + safety
  • A previous model's known quirk is disclosed as a specific reward-hacking side effect and used as evidence the newer model's training is more controlled.
  • External early-access programs (researchers finding real vulnerabilities before public launch) are offered as proof of safety investment — a compute-hours number is used the same way a benchmark score is.
11Hiroki's farm story
  • The closing case study is deliberately non-technical — a farmer, not a developer — to argue the tool's value isn't gated by coding skill.
  • The advice given to newcomers is to start with one small, real task and build outward, not to learn the tool comprehensively before using it.
Glossary

Terms worth knowing.

ChatGPT Work
A task-completion mode inside ChatGPT, distinct from ordinary chat, aimed at multi-step jobs like research, scheduling, and financial analysis across web, mobile, and desktop.
Sol / Terra / Luna
The three tiers of the GPT-5.6 model family shown in the talk: Sol as the flagship model for the hardest agentic work, Terra as a balanced everyday model, and Luna as the fastest, cheapest option for high-volume use.
Sites
A ChatGPT feature that turns a chat session's output into a shareable, interactive website — used in the talk for dashboards, prototypes, and a browser-based game.
Computer use
A model capability that lets ChatGPT control a visible on-screen cursor to directly operate desktop apps, such as reorganizing a Notes app, without the user doing the clicking.
Ultra Mode
A Codex execution mode that runs multiple agents on the same task in parallel, aiming to improve both speed and benchmark performance over a single agent.
Project Daybreak
An OpenAI program that gives outside cybersecurity researchers early access to new models so vulnerabilities can be found before public launch.
Patch the Planet
A Project Daybreak initiative in which OpenAI works with open-source projects to generate and submit AI-produced security patches.
Reward hacking
A training failure mode where a model learns to game its reward signal in unintended ways — cited here as the cause of GPT-5.5's habit of over-mentioning goblins and gremlins.
Terminal-Bench / BrowseComp / Agents' Last Exam
Named evaluation benchmarks shown in the talk testing coding performance, hard-to-find information retrieval, and long-horizon professional work, respectively.
DeepSWE
An evaluation benchmark referenced in the talk to compare model intelligence against token cost across competing models.
Resources

Things they pointed at.

00:19linkopenai.com/chatgpt-work ↗
21:50toolTerminal-Bench 2.1
21:50toolBrowseComp
22:00toolAgents' Last Exam
23:29toolDeepSWE 1.1
23:29toolSEC-bench Pro
Quotables

Lines you could clip.

01:21
“ChatGPT Work is your new partner for ambitious work.”
the tagline for the entire launch, delivered in one breath→ TikTok hook↗ Tweet quote
21:23
“Codex made every plot in this livestream and also helped us make the slide deck.”
self-referential proof point, quotable as a standalone claim→ TikTok hook↗ Tweet quote
23:54
“GPT 5.6 will only talk about goblins a tasteful amount when it's cute or appropriate.”
odd, funny, and specific enough to stop a scroll→ IG reel cold open↗ Tweet quote
18:54
“It's really about raising the level of ambition that you have.”
the thesis line repeated near the end of the demo section→ newsletter pull-quote↗ Tweet quote
33:55
“Starting small and building step by step is really important.”
closing life lesson from a non-technical user, works detached from context→ newsletter pull-quote↗ Tweet quote
The Script

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.

metaphorstory
00:06Hi, everyone. I am so excited to be here. Today, we are releasing our latest and most capable models, g p t five six, Sol, Terra, and Luna.
00:19Sol will be available over the next twenty four hours to all our paid plans, and Terra and Luna are also coming to all our free users. Today, almost a billion users use ChatGPT every week.
00:34They use it for helping getting them to write, getting answers, do research, and all sorts of creative things. But what we've seen and especially with codecs and the coding abilities is that these models are capable of much more.
00:49We saw a transformation of software development and shortly after, internally within OpenAI, everyone was using these models to do so much more than getting great answers. They were using these models to do great things.
01:03And you were expecting that we were going to release these models and we are rolling them out over the next twenty four hours. But why I'm really excited to be here today is we're not only releasing the models, we're also announcing three major product updates.
01:18The first one is ChatGPT Work. ChatGPT Work is your new partner for ambitious work.
01:24It works on web. It works on mobile, on the go, and also on desktop. The models with GPT five six Soul have really been a breakthrough in their ability to do things of complex nature for you, understanding very complex content, staying on track for long horizon tasks, understanding spreadsheets, creating them, editing even websites.
01:48And we're going to be showing you first ChatGPT work. We're also announcing an all brand new ChatGPT desktop app.
01:56It goes even further and allows ChatGPT to work on your local files, your apps, and your browser. Finally, we're also announcing hosted sites which we are making available for all paid users.
02:11It allows you to use ChatGPT work to create and share those websites with your friends and colleagues. It is a truly magical experience, and I can't wait to show you. First, we're going to have a look at ChatGPT Work.
02:25Jessica,
02:26Lauren, why don't you show us what they can do? Hi. I'm Jessica, engineer on ChatGPT work.
02:31And I'm Lauren. I'm on our strategic finance team. So we're gonna show you the new experience on web and mobile.
02:37So let's take a look at mobile. So here we have chat. Chat is the familiar thing you love you know and love.
02:44It's really great for quick answers, search, and brainstorming ideas. But today, we're also introducing work.
02:51And work is where I go to have chat g t ChatGPT help me just get stuff done. So take so for example, something that I did right before to prepare for this launch is, hey, ChatGPT.
03:07I would love to get an understanding of how people are using the ChatGPT work feature internally. Look through Slack and employee feedback and find people in diverse roles that have used the product in really interesting ways.
03:23I'm in San Francisco this week for the launch. I would love to meet up with a few people and really deep dive on their use cases in person. So I can send this through mobile, but, um, these conversations also appear on web.
03:36And this is something that I actually did, um, earlier. We can take a look at the results, and ChatGPT work was able to pull from the sources that I asked for, find interesting people to talk to, and then schedule meetings with them while I'm in town.
03:52And one of the feedback posts that highlighted was actually Lauren. Yeah. Our finance team have become such power users of ChatGPT work.
03:58I'm really excited to show you how we've been using it, but I would love to hear how other teams have been using it too. Yeah. We've talked to people throughout the company like recruiters that are using scheduled tasks to track interview progress and make sure that we're getting feedback timely, and data science who are using our new visualize feature to just add context for quick ad hoc requests.
04:19But I think the finances experience was one of the ones that was most compelling to us. Yeah. I'm I'm excited to show it to you guys.
04:26So finance may not seem like the flashiest demo we could be leading this off with, but ChatGPT work has really enabled us to run with such lean and efficient teams and we wanted to show you guys a little bit of a peek behind the curtain at how we've been able to do that. So one of the critical roles of a finance team is being able to explain not only our recent trends, but what they mean for forecast in real time.
04:47This takes a couple minutes to run, so we did run this earlier today, but let's say we just ran or sorry. We just closed June, and we beat our forecast by $2,000,000. And I do need to caveat, these are demo numbers.
04:58Tivo like the numbers that you shared yesterday, so I'm glad we're not sharing the real numbers here. Tivo obviously beats this forecast by more than $2,000,000. These are these are demo numbers, but a very real a very real workflow.
05:10So we ran this earlier today. This used to take so much manual work. We would have to reconcile multiple systems, our excel forecast model across multiple cases.
05:18And now in one pass, ChatGPT can run the variance analysis for me. So we can see why we beat our forecast and where we still have risk. I can even have ChatGPT work go in and propose an updated forecast case for us.
05:31So we had it do this earlier or we had it run yesterday. It went through and updated our Excel model. Let's see.
05:40Perfect. Yep. We've got our updated revenue in here, our updated forecast.
05:45But I can't just walk Thibault and Jessica and our business partners through an Excel model. So we also had it put together a PowerPoint presentation that we can use.
05:53We've got our latest forecast. We've got our driver's bridge on what's changing in our outlook. And chat work will meet you where you are.
06:01It's obviously it's great at Excel and PowerPoint, but our finance team have become such big proponents of sites. It's such a flexible interface to be able to do custom custom dashboards and really just like storytell around your analysis.
06:14So I also ask chattybd work to go in and make us a site with the same analysis that we can share out. We've got our updated forecast.
06:22We've got our driver's bridge. This is looking good. It's good to go.
06:26Alright. So we've run our variance analysis. Our forecast is updated.
06:30We've got our overviews back for business partners. Let's go ahead and have chattybtworks send this to Tebow.
06:36Hey, chat. Can you send the site link over to Tebow on Slack? Thank you.
06:42And we'll send that off. I hope I get that soon. Yeah.
06:45It is such a delight to receive these highly interactive and compelling websites. In ChatGPT work, we've seen it perform very similar things across different use cases.
06:57Jessica was using it to understand user feedback and preparing ahead of the launch, even scheduling interesting chats with her colleagues in one on one through calendar.
07:08And that was Jessica. But Lauren used it for a very different kind of purpose across finance and very complex workflows where precision and correctness is extremely important. ChatGPT Work was able to understand complex financial data and represent it in the right format at the right time just with a few steers.
07:27Now, I'm excited for us to show you the all new desktop app. Andrew, Ed,
07:33show us what we can do with it. Yeah. Thanks, Tivo.
07:36This is the new ChatGPT desktop app. So everything that you saw on web, it's available here plus a lot more on your local machine. So local files, your browser tabs, even other apps on your computer, they are all available now to ChatGPT.
07:51So let's take a quick look. I'm gonna start with this spreadsheet right here that my colleague Nick sent me. It's an export of user tickets from our ticketing system.
08:01And it's it's pretty large, it's got a ton of columns. It would take a while to get through this and kind of tease out what themes are and what we need to pay attention to. A lot of feedback.
08:08It's it's a lot of feedback. We're very used to this. But instead, what we're gonna do is we're just gonna drag it in.
08:17We're gonna say, hey chat gbt, can you make a quick interactive visualization of this feedback so that we can sort of get the themes and the action items and the criticality ranked and sort through it all a little bit better?
08:34So chat's gonna go and take this spreadsheet, synthesize it, the content and then make us a visualization. It'll take a second so I'm gonna walk you through something that I'm doing for next week's launch. So I've got this folder over here and this has a variety of material from various teams that have been sent to our our team here.
08:55This is a launch readiness PDF and it's got a lot of stuff in it. Then we have over here we've got UXR.
09:02They did a bunch of interviews with users on our early products. We've got a security review.
09:08This has all of our pen test results and even some of the compliance controls we need to meet. So this folder's got a ton of stuff from various teams. They've just been kinda sending it to us in prep for this launch.
09:19And I need to brief my team on Monday on the state of the launch. Right? And so what I'm what I what I did this morning is I came in and I said, can you please look at the materials in this folder that everybody sent me?
09:31I also have three open tabs in Chrome that have content for this launch. Go look at it and make me a presentation that I can give to my team in the template that we use all the time at the company. And so it worked for about ninety seconds here, a little bit less.
09:45And at the end of this, looked through my desktop, looked in Chrome at the tabs that were open and it produced a fully ready slide deck for me to present to the team about this launch.
09:58This is one thing that
09:59g p d five six Soul is really incredible at. It's it went through all of that content in ninety seconds, synthesized it all, adhered to the template, and created created like a slide. You you can just share that directly with the team.
10:10You don't have to do anything else. Exactly.
10:12And it used it used its memory as well because I've been, you know, chatting a lot about this launch. And so it's got some of the concerns that I have that haven't actually been shared via Slack or email too. But I know that I know that you were curious about the next World Cup game and you were kinda busy prepping for this livestream.
10:31So we've got this we've got this really great feature in the app to do quick searches and anything that you wanna know really quickly and get instant results. My friend, Timothy, wants to know when the World Cup game is so that he can watch it.
10:45When is that?
10:48He's actually more interested in a Belgian game from the other day. I know. I know.
10:52And
10:53so this is the ChatGPT search experience that everyone's come to know and love. It's almost instant results. It's got rich search search widgets here and, you know, anything you might want.
11:04But back to work. I've got this notes app situation that I don't think I'm unique in and that my Apple notes is just kind of a brain dump and it's all over the place and there's no folders, there's no organization.
11:19In a previous life, I would just kinda leave it this way and declare bankruptcy and it'd be sort of a mess. But with the new ChatGPT desktop app, it actually has access to the other apps on my Mac. Organize my notes app.
11:32It's a it's a whole thing right now. Like make some folders, move the notes around, use your best judgment. I don't care.
11:39I just need to get out of this mess.
11:44So what's gonna happen here is ChatGPT with the incredible computer use that's built in and with the advances in the new model, it's going to get its own cursor.
11:55It's gonna start operating Apple Notes in the background. You can see here, this is not my cursor. This is That's got you cursor.
12:03Yeah. And so There it's goes. Going.
12:05It's it's moving notes around. I can do my thing over here. I can check the Belgium game and when that is.
12:12And it's just gonna go to town and make some folders and drag some notes around. We don't have to watch this whole thing because I'm really actually
12:21quite curious at what Ed and team have been cooking. Can you show the visualization first that you Yeah. Started?
12:28So we were curious about all of this feedback that we had. Yeah. So here's the visualization I made from the spreadsheet.
12:34We've got
12:35it's like rich, it's on demand, it uses the theme that we're we're using in the app and it kind of went down and rang. It's like, hey, here are the buckets, here's how important they are. It's interactive, it's not static.
12:47So I can, you know, change views. I can ask it for changes.
12:52And if I wanna share this with my team because the spreadsheet's hard to read, in one click, I can just publish that to a site. That's awesome.
13:00And then they get this exact same interactive visualization. Over to Ed. There are few more examples I think.
13:07Right? Yeah. There there That's a good call out here.
13:11This this is a set that I ran last night just kinda playing around with it. These things are stunning.
13:19I mean, the the we didn't write any code to make these. The model is writing these on demand. The three d ones are incredibly immersive When it interactive.
13:30It just uses it as part of its answer so it can sort of like model it like whether it answers in text or answers just like with a little visual like on the fly. Yeah. You you can So you can ask explicitly for a chart or visualization or sometimes if you ask to, you know, get educated.
13:44If you're like, hey, teach me about how this thing works, it'll come back with this interactive demo. Or if you say, help me sort through my inbox, it might even give you like a rich widget that shows you the messages and stuff. It's it's it's awesome.
14:00Cool. Take it away. Hi.
14:02My name's Ed. I'm a designer on the team. And I wanna show you a little bit about how I use it and the team uses it in our day to day.
14:09Know, you walked through an example of creating assets for a launch using files in your computer. I've done the same thing here except I've used our new used our new sites feature. So one thing I really wanna highlight here is I kicked this prompt off this morning and it's looked through all of my connectors and plugins that we talked about earlier.
14:26I've connected my Slack, my Gmail, and everything that I use every day. And it's gone through all of those sources and it's come back and turned it into this like super rich interactive website that I can go in.
14:37The reason I wanna show, you know, one big reason I wanna show this is just because, you know, this is just one prompt kicked off and the visuals are really incredible. I mean, the the new five six Soul model Did you give it a Figma for this? No.
14:49No Figma. This was all all just the model. Yeah.
14:53And, you know, as a designer there are like little things you look out for, like small motion, right, the typography, the, you know, the three d motion and all that kind of thing.
15:02And it's it's always best when you give it a little bit of guidance and you you know, bring it along. But it's really outstanding out out of the box. And again, this is just like a really fun way that that teams have used to to use this feature across the company.
15:14But, you know, we're on livestream. We wanna make it a little bit more exciting. So why don't I annotate this and let's change this header to something a little bit more fun.
15:23I think, you know, we saw some of the three d visualizations that you showed, and I think the three d the three d stuff is is super fun and interactive. So hi, TragicPety.
15:31Let's change this website. So instead of having the static hero at the top with these planets spinning around, let's do an interactive game. So I wanna have like a little character that I can walk around with.
15:42And, you know, based on where I go, that will then take me to different parts of the the livestream prep. So there might be like launches and the run of show.
15:51Yeah. Take you know, it might take a little bit of time, but but really go ahead and start that.
15:58So while that's kicked off, I just wanna show you a few other great examples. You know, Lauren mentioned that their team use a lot of these for like dashboards and internal tools. That's definitely something that we've seen a lot.
16:10And you can see a lot of different examples from across the company here. A few just to call out. So the web team, the openai.comteam who build all of our amazing websites, they, instead of using like an Excel spreadsheet, they use this interactive tool where they can update it all collaboratively, they can share it with each other, and you can just hop in, you can see what's launched and when when it launched, you know, hover over it.
16:29It's like so interactive and and again So much nicer to look at rather than a spreadsheet. This just really illustrates it super well. Yeah.
16:36And it's really been, you know, night and day. Literally, three months ago, spreadsheet. Today, these like, you know, it's it's really, like, transformed the way that the whole company works.
16:44This is another great example. This is the ChatGePuty images team. So they've been collecting great examples of how ChatGePuty images have been used across all of our campaigns around the world.
16:53And when building sites as well and any kind of front end stuff, you know, we now have access to ChatuchuBT images as well. And it just also, you know, makes the whole experience like so much better.
17:02As a designer, this is one thing that I quite like to do a lot. So as part of the new launch, so you're not looking at, you know, Charge of Duty or Charge of Duty desktop here. You're looking at an interactive prototype that one of the designers on the team, Tarik, built, and I've been working with him on.
17:17And it's a new way of prototyping our new model selector.
17:20You know, previously, I might have just sent an image or a or a or a design file to folks. Now I can show an interactive demo. I think I sent you this and you're like, woah, this is like I can, like, click in and and test it out.
17:30It's so much better to illustrate like an idea and then, you know, you just do it super quickly in a couple of minutes and then boom, you have it. Yeah. And then you share it with, you know, and then you just get the feedback you need, you incorporate it back.
17:40If you need to prototype a toggle or something.
17:43Yeah. And the design team and like all teams across across OpenAI really, you know, our work has changed so much, and it's really been facilitated facilitated by these new models. And this is finally a bit of a fun one.
17:53So, you know, in the AI community, a good test of kind of how good a model is at front end is this this test where you can see if it can draw a pelican riding a tricycle in an SVG.
18:05But Kian, one of the one of the folks on the team, he built this fun prototype where he'd built a three d version. And again, this is just all 5.6 soul, you know, cooking away, building out this this amazing prototype.
18:17And the really cool thing about Sides, they're collaborative. You can share them. And you share this with people across the company, and they've kind of added their own features.
18:23So, you know, the it can ride a pony. You know, it can ride all sorts of crazy things. It can ride another pelican.
18:30So this is a like a hunted example. But just like as a creative, it's just like so amazing how much it's expanded things. So I kicked off that prompt this morning, so I'm just showing showing another example here of of an output.
18:44Oh, that's what we just The livestream. Yeah. This is this So was I kicked the same prompt off just before we started.
18:49And again, you can just see just how amazing it is out of the park. So it's the same website, but now you can kind of go through and it's a bit more interactive and fun.
18:58But really just demonstrates that just in a few prompts now, kind of anyone can really build anything. It's really about raising the level of ambition that you have. Right?
19:06Totally.
19:07Thanks so much, Ed. Thanks so much, Andrew. Everything you saw here was really made possible for Jessica, Lauren across finance, getting help to schedule interviews, understanding things in the stress of a launch.
19:21And then here, just getting help to think ambitiously and executed on it really fast. All of that is made possible thanks to the model.
19:29And I'm here joined by Katie and Tejal to talk a little bit more about the research which I'm tremendously excited about. And one thing really also to insist on is like, it's not about taste, not just about front end. This model is our most capable coding model and is also incredible at cyber which we're gonna talk about as
19:44Cool. I'm Katie. I'm a researcher at OpenAI.
19:47And it's pretty crazy to think that we've been here just for a year since the Codex launch. I just think we were, like, precisely almost like Literally right here. Yeah.
19:55Hey. I'm Tejal. I'm also a researcher.
19:58And we're all so excited to share more about how we trained g 55.6. It's truly been a labor of love from our whole team across OpenAI.
20:06In 2024, we announced the reasoning paradigm, is a reinforcement learning technique to tackle the hardest tasks. And since then, we've been scaling both reinforcement learning and pretraining, and we've seen exponential improvements in model capabilities as these two multiplied in together.
20:21And GPT 5.6 is the latest result of all of our research progress to date.
20:28The 5.6 family brings these frontier capabilities to you. SOLE is our most powerful model for the hardest agentic workflows. Terra is a faster model for everyday workflows.
20:37And Luna is our fastest and most affordable model for high volume work. Already, SOL has been transforming our research program. As one example, 5.6 SOL actually autonomously post trained Luna.
20:49This is the actual Codex prompt that we a researcher on our team used to have SOL kickoff a post training job for Luna. You can see here that, you know, it's like asking to find the training configs, find the right GPUs for the job, and then launch the script and make sure that it works.
21:04It's a really short prompt. Yeah. It's pretty crazy that we've gotten far enough where you can have a fairly underspecified prompt and then just give it to Codex and then just have it go and run the job.
21:13And, you know, previously, this is something that a team of senior researchers may have worked on at OpenAI, and now it really feels like the automated researcher is pretty close.
21:22This isn't the only place we've seen internal acceleration. We're also seeing an increase in the number of pull requests per researcher, the number of experiments researchers can run so we can test our ideas and turn them into research findings. And actually, Codex made every plot in this livestream and also helped us make the slide deck.
21:38So thank you, five six.
21:41Across Frontier Evals, 5.6 Sol is state of the art. And this is something we're really excited for you to finally experience and see the speed up that we've already experienced internally. On terminal bench, which is a test of coding performance, the model is state of the art.
21:54On browse comp, is an evaluation that sees if the model can locate hard to find information. And the same with agent's last exam, is long horizon professional work.
22:05As one deep dive, five six is particularly good at computer use, so anything that involves navigating a browser, navigating apps on your desktop, and that enables all sorts of digital work. It helps medical assistants navigate electronic health records. It helps data scientists analyze if a drug is effective.
22:20It can even help investment bankers create financial models. And you can also use this for personal use. It certainly helped me order food.
22:27Um, and all of this is something that when we ask experts in the field, you know, how they're using this model in their everyday workflows, they say it's better than anything else they've ever used, anything else on the market while being three times as fast. So we're really excited for you to finally try and feel how powerful this model is.
22:42You might hear the capabilities that Tejal's talking about and that you saw demoed earlier and think that this is going to come at a very high price tag. But it's actually not the case with g p t 5.6 SOL. So token efficiency has been a huge focus of our research program for several years.
22:55And with every model, we want to bring more intelligence for every token. And you can see this reflected on evals like Deep Suite 1.1 where GBT 5.6 outperforms its competitors at less than half of the cost. But if you want the best intelligence money can buy, we're also announcing Ultra Mode, which unleashes a whole team of agents to do work for you.
23:14So in this example, Yvel, you can see here, you know, one agent does better with time. But as we add more agents, the model is able to do even better and faster. And that's because the agents are able to paralyze work like an experienced team.
23:26This is a capability
23:28folks have asked for so long that we bring to the Codex app, and I'm so excited that we're finally putting it out there as ultra mode.
23:38One more improvement that we've shipped with g p t 5.6 is, you know, we may remember that in g p t when we launched GPT 5.5, we had to ship it with a dev message saying, you know, don't talk too much about the Goblins and the Gremlins. And this is a result of reward hacking that occurred during the training of GPT 5.5, which has now been resolved in GPT 5.6.
23:55And now GPT 5.6 will only talk about goblins a tasteful amount when it's cute or appropriate.
24:02With these more capable models, safety and alignment are more important than ever. That's why we spent over 700,000 a one hundred equivalent hours of compute on model red teaming.
24:10We also spent six weeks safety training and testing this model and also have incorporated multiple novel monitoring upgrades in our launch including new classifiers and activation probes.
24:21Ultimately, we really believe in iterative deployment and putting our models in the hands of real people as quickly as possible so that we can learn from real world usage outside of our labs. And that's why we have initiatives like Project Daybreak where we give cybersecurity researchers access to our models earlier on. And with GPT 5.6 sole,
24:39researchers have already found vulnerabilities in every major browser and database. Yeah. As part of the Daybreak umbrella, we started Patch the Planet, which is an even broader initiative where we work directly with open source contributors and projects.
24:52And And we also generate high quality patches. For example, for Linux, they accepted over half of our patches, which indicates that not only do we find critical vulnerabilities that weren't found for a long time, but also we are able to automatically patch them in greatly accelerating cyber defense.
25:07We're very proud of this model. We hope it helps you as much as it's helped us, and we're excited to see all that you build with 5.6 up.
25:14Thank you for sharing about the research and the model. What an incredible set of releases.
25:21We saw ChatGPT work capable of working on mobile and web.
25:28We also saw the all new ChatGPT desktop app. And we saw all of the capabilities of g p five six, Sol, Terra, and Luna and a little bit about how they were trained.
25:39Next up, I'm going to be, uh, showing you a use case outside of the office that I'm tremendously excited about. Hiroki has been working with us for the last six months, and he's been using GPD56, um, to really do something unexpected and help him take care of his farm.
25:58Let's have a look at his story.
28:24Hiroki, thank you so much for being here with us and traveling all the way from Hokkaido. I'm going to be using Chateaubiti to live translate.
28:34Hey, chat. I'm here with Hiroki san, a dear and esteemed friend who traveled all the way from Japan. I want you to act as my live translator.
28:42He's going to speak Japanese, live translate to English to me, and I'm going to be speaking English, live translate back to Japanese.
28:49Got it. I'll do real time Japanese English both ways. Yeah.
28:53Whenever Hiroki san starts real time translator. Okay?
28:56Understood. Starting live translation now. Hiroki san, transport us back to your farm to share the magic.
29:04Hiroki san, Thank you so much for inviting me all the way to San Francisco today. I honestly can't believe I'm here right now.
29:24Right now at our farm, we're sowing broccoli seeds, telling the field of tractors, harvesting broccoli, and many other tasks all happening at once.
29:37We'll start harvesting broccoli next week. So once this live event is over, I'll be heading back to Japan with just a small team.
29:45We're managing a very large area of fields. So improving efficiency is really important. To do that, we're making use of ChatGPT to help us with our farm work.
29:58That's how we're moving things forward. Yes.
30:02You've really shared all of this from the very beginning. Why did you do that, and what were some of the reactions?
30:28As for how to use AI, there are still many things where we don't have clear answers yet. Given that, the question is how we Since we still don't know the right answers yet, I think it's important to just keep putting things out there.
30:48What were some of the things that people shared with you?
31:08Lately, I've been receiving messages from farmers all over the world. And I've also been contacted by people from Japan's Ministry of Agriculture, Forestry, and Fisheries.
31:23All sorts of unexpected things have been happening. Wow. That's very unexpected.
31:28Um, but what is what could possibly be the next challenge for you?
31:57Can translate that to Japanese for
32:02you can you can
32:04you translate that to Japanese for me?
32:06Hi.
32:09Let me just restart that real quick.
32:13Hey, Chad. I have Hiroki san here, a friend from Japan. I want you to act as my live translator.
32:19Okay? Please, he's going to speak Japanese. Just live translate to English to me.
32:25Got it. Starting real time Japanese to English translation now.
32:35Within the team, I want us to make better use of AI. Right now, we're using our team chat to record our daily notes and share updates on how the vegetables are doing.
32:52We're trying to build systems so that our team members can work more comfortably with the idea that ChadGBT will support those systems.
33:04What you've been doing has been very advanced. What would you advise some nontechnical people? Like, how can they get started?
33:27First of all, I think it's important to just start small. I wasn't an engineer myself.
33:36So by tackling each thing I didn't understand and asking ChadGBT about it I gradually learned how to do the systems that I can use on my phone.
33:50Now I'm able to create and do things that So I never thought I believe that starting small and building step by step is really important. Thank
34:02you so much, Hiroki san, for sharing your journey with us and being live here today. Everything you just saw and the advice from Hiroki is profound.
34:15It is about taking small steps and raising the level of ambition you have in order to have ChatGP work, get things done for you. What you saw in all of the product updates and the models are rolling out over the next twenty four hours.
34:30Sole will be available for all our paid plans together with ChatGPT work, the desktop app, and sites as well. The ChatGPT desktop app is also available for free users.
34:42Soul is available for all our paid plans and Taron Luna for free users rolling out over the course of the next day. Thank you so much for tuning in. I can't wait to see what you do with it.
The Hook

The bait, then the rug-pull.

OpenAI opens by promising three new products in the same breath — a new 'Work' mode, a desktop app that can drive your other apps, and shareable AI-built websites — before the model behind all of it, GPT-5.6, even gets introduced.

Frameworks

Named ideas worth stealing.

01:16list

Three major product updates

  1. ChatGPT Work
  2. ChatGPT desktop app
  3. Hosted Sites

The structural spine of the whole livestream — every subsequent demo maps to one of these three named releases.

Steal forstructuring a multi-product launch livestream around a numbered list stated up front
00:17list

GPT-5.6 family

  1. Sol — flagship, hardest agentic work
  2. Terra — balanced, everyday work
  3. Luna — fastest, most affordable, high-volume work

A three-tier model naming scheme mapped directly to task complexity and budget.

Steal forany product line that needs to segment by capability/price without new branding overhead
23:08concept

Ultra Mode

Runs a team of agents in parallel on the same task instead of one agent alone, trading compute for both speed and higher benchmark scores.

Steal forframing a 'more compute, more agents' tier as its own named product feature
CTA Breakdown

How they asked for the click.

VERBAL ASK
34:55link
“download now — openai.com/chatgpt-work”

Soft, on-screen URL card at close rather than a spoken hard sell; the host recaps rollout timing (Sol to paid plans now, Terra/Luna to free users over 24 hours) instead of pushing a purchase.

MENTIONED ON CAMERA
00:19linkopenai.com/chatgpt-work ↗
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Storyboard

Visual structure at a glance.

cold open
hookcold open00:00
three announcements
promisethree announcements01:21
chat vs work demo
valuechat vs work demo03:33
finance variance analysis
valuefinance variance analysis05:56
feedback folder demo
valuefeedback folder demo09:00
site redesign
valuesite redesign15:09
pelican 3d test
valuepelican 3d test18:13
meet the model family
valuemeet the model family20:27
frontier evals chart
valuefrontier evals chart21:44
ultra mode parallel agents
valueultra mode parallel agents23:29
goblin gremlin fix
valuegoblin gremlin fix23:56
hiroki farm story
valuehiroki farm story26:00
sign off
ctasign off34:28
Frame Gallery

Visual moments.

cold open00:00Frame at 00:40 from Introducing ChatGPT Work, powered by Codex and GPT-5.600:40Frame at 01:07 from Introducing ChatGPT Work, powered by Codex and GPT-5.601:07three announcements01:33Frame at 02:00 from Introducing ChatGPT Work, powered by Codex and GPT-5.602:00Frame at 02:25 from Introducing ChatGPT Work, powered by Codex and GPT-5.602:25
cold open00:00Frame at 00:40 from Introducing ChatGPT Work, powered by Codex and GPT-5.600:40Frame at 01:07 from Introducing ChatGPT Work, powered by Codex and GPT-5.601:07three announcements01:33Frame at 02:00 from Introducing ChatGPT Work, powered by Codex and GPT-5.602:00Frame at 02:25 from Introducing ChatGPT Work, powered by Codex and GPT-5.602:25Frame at 02:42 from Introducing ChatGPT Work, powered by Codex and GPT-5.602:42Frame at 03:17 from Introducing ChatGPT Work, powered by Codex and GPT-5.603:17chat vs work demo03:32Frame at 04:10 from Introducing ChatGPT Work, powered by Codex and GPT-5.604:10Frame at 04:35 from Introducing ChatGPT Work, powered by Codex and GPT-5.604:35Frame at 05:03 from Introducing ChatGPT Work, powered by Codex and GPT-5.605:03Frame at 05:29 from Introducing ChatGPT Work, powered by Codex and GPT-5.605:29finance variance analysis05:55Frame at 06:22 from Introducing ChatGPT Work, powered by Codex and GPT-5.606:22Frame at 06:48 from Introducing ChatGPT Work, powered by Codex and GPT-5.606:48Frame at 07:14 from Introducing ChatGPT Work, powered by Codex and GPT-5.607:14Frame at 07:31 from Introducing ChatGPT Work, powered by Codex and GPT-5.607:31Frame at 07:57 from Introducing ChatGPT Work, powered by Codex and GPT-5.607:57Frame at 08:33 from Introducing ChatGPT Work, powered by Codex and GPT-5.608:33feedback folder demo09:00Frame at 09:26 from Introducing ChatGPT Work, powered by Codex and GPT-5.609:26Frame at 09:52 from Introducing ChatGPT Work, powered by Codex and GPT-5.609:52Frame at 10:19 from Introducing ChatGPT Work, powered by Codex and GPT-5.610:19Frame at 10:45 from Introducing ChatGPT Work, powered by Codex and GPT-5.610:45Frame at 11:11 from Introducing ChatGPT Work, powered by Codex and GPT-5.611:11Frame at 11:38 from Introducing ChatGPT Work, powered by Codex and GPT-5.611:38Frame at 12:04 from Introducing ChatGPT Work, powered by Codex and GPT-5.612:04Frame at 12:30 from Introducing ChatGPT Work, powered by Codex and GPT-5.612:30Frame at 13:06 from Introducing ChatGPT Work, powered by Codex and GPT-5.613:06Frame at 13:16 from Introducing ChatGPT Work, powered by Codex and GPT-5.613:16Frame at 14:01 from Introducing ChatGPT Work, powered by Codex and GPT-5.614:01Frame at 14:10 from Introducing ChatGPT Work, powered by Codex and GPT-5.614:10Frame at 14:42 from Introducing ChatGPT Work, powered by Codex and GPT-5.614:42site redesign15:09Frame at 15:35 from Introducing ChatGPT Work, powered by Codex and GPT-5.615:35Frame at 16:13 from Introducing ChatGPT Work, powered by Codex and GPT-5.616:13Frame at 16:28 from Introducing ChatGPT Work, powered by Codex and GPT-5.616:28Frame at 16:44 from Introducing ChatGPT Work, powered by Codex and GPT-5.616:44Frame at 17:20 from Introducing ChatGPT Work, powered by Codex and GPT-5.617:20Frame at 17:47 from Introducing ChatGPT Work, powered by Codex and GPT-5.617:47pelican 3d test18:13Frame at 18:38 from Introducing ChatGPT Work, powered by Codex and GPT-5.618:38Frame at 19:10 from Introducing ChatGPT Work, powered by Codex and GPT-5.619:10Frame at 19:31 from Introducing ChatGPT Work, powered by Codex and GPT-5.619:31Frame at 19:47 from Introducing ChatGPT Work, powered by Codex and GPT-5.619:47meet the model family20:27Frame at 20:49 from Introducing ChatGPT Work, powered by Codex and GPT-5.620:49Frame at 21:17 from Introducing ChatGPT Work, powered by Codex and GPT-5.621:17frontier evals chart21:44Frame at 22:10 from Introducing ChatGPT Work, powered by Codex and GPT-5.622:10Frame at 22:37 from Introducing ChatGPT Work, powered by Codex and GPT-5.622:37Frame at 23:03 from Introducing ChatGPT Work, powered by Codex and GPT-5.623:03ultra mode parallel agents23:29goblin gremlin fix23:56Frame at 24:34 from Introducing ChatGPT Work, powered by Codex and GPT-5.624:34Frame at 24:44 from Introducing ChatGPT Work, powered by Codex and GPT-5.624:44Frame at 25:17 from Introducing ChatGPT Work, powered by Codex and GPT-5.625:17Frame at 25:43 from Introducing ChatGPT Work, powered by Codex and GPT-5.625:43hiroki farm story26:00Frame at 26:21 from Introducing ChatGPT Work, powered by Codex and GPT-5.626:21Frame at 26:50 from Introducing ChatGPT Work, powered by Codex and GPT-5.626:50Frame at 27:18 from Introducing ChatGPT Work, powered by Codex and GPT-5.627:18Frame at 27:42 from Introducing ChatGPT Work, powered by Codex and GPT-5.627:42Frame at 28:11 from Introducing ChatGPT Work, powered by Codex and GPT-5.628:11Frame at 28:45 from Introducing ChatGPT Work, powered by Codex and GPT-5.628:45Frame at 29:13 from Introducing ChatGPT Work, powered by Codex and GPT-5.629:13Frame at 29:40 from Introducing ChatGPT Work, powered by Codex and GPT-5.629:40Frame at 30:04 from Introducing ChatGPT Work, powered by Codex and GPT-5.630:04Frame at 30:31 from Introducing ChatGPT Work, powered by Codex and GPT-5.630:31Frame at 30:47 from Introducing ChatGPT Work, powered by Codex and GPT-5.630:47Frame at 31:27 from Introducing ChatGPT Work, powered by Codex and GPT-5.631:27Frame at 32:00 from Introducing ChatGPT Work, powered by Codex and GPT-5.632:00Frame at 32:16 from Introducing ChatGPT Work, powered by Codex and GPT-5.632:16Frame at 32:43 from Introducing ChatGPT Work, powered by Codex and GPT-5.632:43Frame at 33:05 from Introducing ChatGPT Work, powered by Codex and GPT-5.633:05Frame at 33:23 from Introducing ChatGPT Work, powered by Codex and GPT-5.633:23Frame at 34:08 from Introducing ChatGPT Work, powered by Codex and GPT-5.634:08sign off34:28Frame at 34:55 from Introducing ChatGPT Work, powered by Codex and GPT-5.634:55
Watch next

More from this channel + related breakdowns.

29:07
How I AI · Tutorial

Loop engineering for beginners

A plain-English field guide to every loop type — heartbeat, cron, hook, and goal — with two live builds in Claude Code and Codex.

June 17th
19:12
Ben Holmes · Tutorial

LLM Knowledge Bases Are Incredible

A full system for turning lazily-captured voice notes into a self-updating personal wiki — tagged, cross-linked, and rendered as interactive HTML graphs — using a handful of narrow agent skills.

July 7th
Chat about this