Name: Administering Work in ChatGPT
Uploaded: 2026-07-15T21:18:48.169Z
Duration: 45 min 18 s
Description: Administering Work in ChatGPT

Transcript for "Administering Work in ChatGPT": Hey, everyone. How's it going? I'm Alexander. I'm on the product team here. And I'm Hannah, and I'm on the product team as well focusing on enterprise. Is this your first webinar? This is my first webinar I'm doing. It is. I'm feeling a little awkward about it, but hopefully, you guys enjoy. Yeah. So we're here to talk to you today about Chat. G work. It's both of our first webinars, so please, light up the chat, with chat and questions, and, we'll be answering kind of as we go. Yeah. I'm really excited to share with you today. Cool. Okay. So, as you may have seen, we launched ChatGPT, and GPT 5.6, a new family of models last week. And we're super excited to share it with you. Some context about us is we've been really focused on building out our agents at OpenAI for the past year, and our primary focus now is enterprise. So we are really excited. Let's go to the next slide. Just really excited, to dig in with you guys. You know, this was this webinar was just like an idea that we were brainstorming as we were as we were thinking about shipping the feature. And so, you know, we definitely, love as much engagement from you all as we go and would love feedback at the end on if this is a thing we should consider doing more often. We're thinking maybe even as often as monthly if there's demand. Yeah. And I think the thing to note is, like, we know it's getting increasingly more complex for you guys, and so that's why we wanna make sure that we have a steady cadence of communicating, answering questions, and knowing how you guys can really control and configure the environment so that it meets your policies. Yeah. So hopefully, at the end of this, you have a quick, quick understanding of both the new model, and, also the new product work. You feel empowered to configure it for your team, maybe give it a test ride. Feel free to test as we go, in fact, and then you're off to the races. So a little bit about what we'll talk about. We're gonna talk about, the launches last week. Hannah's gonna do a demo of, hopefully, all the stuff that you're gonna need, to configure and understand what's going on. And then we'll talk a little bit about some sort of recommendations for how to manage the change at your companies, and then, we'll just be open for q and a at the end. Perfect. Alright. Cool. Well, kicking us off, we wanted to first start with introducing GPT 5.6. The reason why we wanna start here is that GPT 5.6 and our focus on bringing you increased intelligence is what shapes all of our product strategy that to come afterwards. And so, the way that we think about GPT 5.6 is it's not only our most intelligent model, but it's also specifically designed for a few different things. The first is that it is more persistent than any of our models in the past, meaning that it can handle steps, tools, and decisions more autonomously, more persistently to get work needed done without needing to ask for guidance. The other is that it's extremely context efficient as well. So it's able to find the right context for a given task with fewer retries, meaning that it can be able to do what you want, find what you want, work for longer to give you the best results without needing to be able to, for you to go in and edit, and to be able to point it at different sources. It should give you the best answer on the first try. And so what that means is when you wanna create things like artifacts of spreadsheets or documents, you can be able to do this both quicker, more token efficiently, and ultimately, you're getting better outcomes. So when we think about looking at at that as how it stacks across our data, thing what we're most proud about right now is GPT 5.6 leads across our Frontier Evals, but it leads across Frontier Evals not just within the coding domain, but also how to be able to use the computer more broadly and being able to do larger tasks and knowledge work more generally. And so you can see at the bright blue line here where GPT 5.6 stacks up, it really wins across evals across the full life cycle of things that you'd wanna use model intelligence for. And in addition to that, it's also our most token efficient model to date. And so as your usage of these models increases across your organization, we know that the thing that is top of mind for all of you is how do you be able to harness this intelligence while also being able to control your costs. And the way that we're focused on delivering that is through token efficiency. And so we think the graph here is super interesting just to set context. So across the bottom axis here for an average task in DeepSue benchmark, what we find is that if cost is on the horizontal axis here and on the vertical axis is, the performance or the accuracy of that task. What we're finding with GPT 5.6 sole on max effort is that for an average task, it costs about a third as less. So it's $8 to perform that task while it's able to produce the same intelligence, if not better than the models of 75%. So it's achieving higher accuracy with third of the cost relative to these other models, which is something that we know is really important as it translates into cost for you and how do you think about managing your budgets. Yeah. When we're talking, to, to execs and admins with our customers, we're always very happy whenever, you folks bring up cost because our models are just great at it. Super scalable. Absolutely. Not only that, we're also giving you as part of this model launch a few different ways to interact with GPT 5.6 intelligence, and so we've packaged it in three different model families. The first being SOL, so you can leverage GPT 5.6 SOL. For some of your most ambitious work, it's something that it brings our highest intelligence. We also package it into Terra, which is more balanced. So it has the intelligence, but is better suited for more, like, efficient and everyday work. And then we also provide you the most cost effective version of this, which is it's gonna be both fast and it's designed more for high volume work. So we give you the capability of deciding which model that you want to use at any given time depending on the work loads that you have. So with that, I would actually love to to hand it over to Alex to talk about GPT work. Cool. So, by the way, thanks everyone for the chat and the q and a. Hannah, are you a real person? I am a real person. That's what a real person would say. Is this AI? Yeah. I'll leave it up to you all to decide if I'm a real person. I don't even what is real? I don't know. But, okay. And then there's some other good questions in here as well. I will talk about it a little bit now about the features, you know, between the different parties products. And let's see. Yeah. We're just collecting questions for q and a. A lot of the questions are around, like, hey. How do I turn this thing on? You know, how do we restrict things? So I think we'll answer a lot of those during the demo, and then we'll make sure to loop back to those at the end if we aren't answering them. So please please keep popping in, in those questions. I'm just kind of taking a scroll through now. Okay. So let's talk about Workovit, and then, I'll talk a little bit about the different scene of features. And, actually, one ask is I'm just scrolling here. Boom. Boom. Boom. Who was asking this? Someone was oh, yeah. America Olivia was asking about, chat the new chatbot codec doesn't have all the features that chatbot classic has. And are we planning to bring, consolidate the interfaces? The answer is yes. But if anyone has specific features from ChatGPT chat, that they're really interested in seeing brought over sooner, please drop those in the chat. It would be very interesting for us to read and think about those. Okay. So, actually, let's go back one slide if that's okay. Oh, yeah. I guess that's just that. So, ChatGPT work. For those of you a lot of you have probably seen this already, but we have been working on a really powerful coding agent. It's called codex. It's the most powerful coding agent out there. And if a sort of fun story here is that, earlier this year when we first launched the desktop app for codex, we started to see that developers were using codecs for much more than just writing code. They were using codecs for everything across the software development life cycle, whether that was, you know, like thinking about what feedback to work on, making plans, writing code, and then, like, working, to deploy that code, you know, monitoring production traffic and, like, collecting more feedback and so forth. And so we noticed that, actually, codecs was useful for much more than just coding. And, a month or two later, we said, hey. Like, how what would it look like if we actually made codecs great, for everyone, not just for software developers, but for anyone trying to do work at your company. And, ultimately, that journey is what led us here to ChatGPT work. So, basically, you have the full power of our agent, but packaged in ChatGPT, so it is now the most powerful version of ChatGPT available to all your professionals. So let's go to the next slide here. I can just do this. Right? Cool. Okay. So I'll talk a little bit about what you have now. So our vision for you and your teams is that you don't have to think about different apps and different things to download. You can just download one thing, and that's just the ChatSpiti app. And similarly, if you're on web or you're on mobile, you just go to chatsp.com or you just download get the ChatSpiti app there, respectively. And there is chat. Chat is a thing that you know and love. You're used to it. It's for everyone. It's fast and conversational, and everyone knows how to use it already. Now with Chatfetchd work, you get an agent, and agents are much more powerful. And so they're really are offering for professional use. If you don't just wanna ask questions, but you want something that can go and understand, like, a complicated task starting from, hey, like, understanding, researching within your company's tools, like, information. What should I do? What would a good plan be? What do we need to know? And then actually producing a deliverable. So not just giving you an answer, but actually, like, turning that into a deck, turning that into analysis, or maybe even going into a tool that you use and making updates in that tool. That's what agents are for. So in short, they're for, like, delegating longer, harder, multi step work and doing so connected with your tools and your enterprise systems. So, you know, chat is something that you can talk to and, you know, do a q and a with. An agent is something that you can embed the way that your team works to take on hard tasks. And so that is what you should use Chat 50 work for. And in the meantime, we still have codecs. Codecs is our agent for developers and technical users. So, basically, I would say if if I was you and you're talking to your team and you're talking to, like, highly technical users or developers, they'll find their way to codecs or you can send them there. And for anyone who doesn't wanna be reading code and wants to just get, like, knowledge work done, send them to ChatGPT work. So let's go to the next slide. Cool. So I've been using this term agent, so I thought I would just, like, quickly talk about what an agent is. And, like, broadly, the idea is an agent, like I was saying, doesn't just answer your questions, but it can take your ideas and turn those into finished work. So if we go from left to right here, we can kind of look at, like, what is the flow of how you might work with one. So this is pretty basic, so bear with me. But to start, you trigger the agent. For most of us, you know, we'll start out by triggering the agent by prompting it. So you might ask a question. Right? You might ask, hey. Like, when is the webinar planned? Help me prepare content for this webinar. Help me brainstorm demo ideas and then help me create my demo. That might be your prompt when you trigger. However, once you have things that you know work, for example, I use my agent very often to do pretty simple things like draft replies to messages. Then you might want that to be done automatically without you having to ask the agent to do it. And so, we can share more about this later if there's interest. You can also set up automations for your agent. And so once you have a thing that works, now you can have it work automatically. So once you've triggered the agent, then the first thing that it'll usually do is it'll go and gather context. And this is just like, you know, a human teammate. If you give them a task, the first thing they'll do is they'll go research. So, it'll go and gather context from your computer if you have the desktop app or via plug ins. And I see some questions about those. We'll answer those later. Then it'll go perform your task. It'll reason about your task and plan. And then finally, it'll produce an artifact for you. So if you say, hey. You know, I'm interested. Normally, I would have to produce a deck or would have produce a spreadsheet for this. You can just ask the agent to do exactly that same thing, maybe even provide it a template, of, like, how you normally do that work at the company, and it'll produce something that fits in with your how your work. Next slide. Chat 50 work is available everywhere that your team works. So you can go to chat50.com and use it. You can download the desktop app, which is the most powerful place to use it. And, actually, you can also use it on your phone, and I use it primarily personally, but on desktop and mobile. Same for you? Or Yeah. Same for me. Cool. Okay. And I know I've been talking for a while, so we won't spend too long here. But just so you all know, ChatKey work is going through our standard enterprise preview process. So today, it is off by default for all of your teams. We'll talk about how to turn it on. We'll answer your questions. And then on July 23, unless you opt out, it'll turn on for your entire teams. Next slide. Cool. And Hannah's gonna demo more about this, but I just wanted to briefly touch on before we get into the demo how we think about control and governance for agents. And, basically, what we have is we have at every level and altitude of where you might want to govern the agent an option free. So starting at the bottom at the most strictest level, you as an admin can define what are the requirements for how I need my team to use the agent. And these are things that you can set You can say, hey. By default, I want my company to work in this way, but I also trust that my team can decide to work in a different way if they want. The next level above that is user defaults. So the user can just decide how they want things to work normally. And then lastly, while the agent is actually doing work, users can decide, to overrule, like, maybe, you know, for the specific task, I want to use a smarter reasoning level, and the agent can ask them questions like, hey. Can I run this? You know, can I, send this email for you? And the user can approve that. So we have, like, all the way from, like, what you define to be strictly true to what you define to be default up to the user level config. So, Hannah, we'll share a little bit more. Yeah. No. That was a great overview. You guys are probably wondering what it looks like, so why don't we go ahead and jump into demo? Alright. So what I'm gonna start here first is what you're seeing is it's you're seeing our formerly codex app now renamed as the ChatGPT desktop app. The desktop app is the primary surface that we use internally within OpenAI. We think that this is the most powerful surface and the surface that we'd actually encourage you guys to be able to do security reviews to roll out more broadly. This is because it can interact with any of your local files or folders and and any plugins that you provision access to. And so what you're seeing within this desktop app is it should look pretty familiar with you. We have a main composer here just as what you would have within ChatGPT. But the big thing that's new as part of this launch is that we've introduced this high level toggle up here. And so the toggle previously, it was only the codecs app, and so you had access to codecs. And what we're finding as Alexander was mentioning is that a lot of the users of the codecs app was moving from just your classic developers that were leveraging it to being able to do coding actions to leveraging the codex app for knowledge workers. And so as part of that, that's really been the genesis of us introducing this ChatGPT work one. And when you switch over to ChatGPT work, you're getting the same power underneath the hood of the entire codex harness and its agentic capabilities, but we're packaging it in a way that's much more friendly and usable for your broader organization. Everything behind the scenes, though, remains the same if we're across a way in which, the agentic action is reasoning through the harness both within codecs and ChatGPT. This is a way for just more of your organization to be able to harness that power. And so that's the that's one of the biggest switches here is it's really just introducing a more accessible interface for ChatGPT work. And then within this, you have the ability to pin certain actions. You have the same project construct that you would have within, ChatGPT, and then you also have the ability to just kick off tasks as well. So I'm gonna just show you a few different things, that you can use this for. So first is as I was preparing for this, I something that you could do within ChatGPT, but just showing it how you could do here too. I sent a picture of the what you'd seen in that screenshot, and I said, hey. Take a look at this and just remind me what each of these eval benchmarks stand for. And so it went and it actually worked for about thirty seconds. You can see here it searched the web and it came back with a very clear response of what each of these different evals are and what they stand for, which helped me to prepare for this presentation. Pro tip. You don't need to spell correctly when you're talking to your agent. You can just you can miss letters and the agent understands what you mean. That is a prototype. You can see I ran out of time before preparing for this. As you wanted to do something that's more, like, complex to showcase the power of what you could do with ChatGPT work, another thing that I did in preparation for this is, I actually asked to be able to look at the usage data, and I didn't even specify where this usage data was. I said, like, you know, look at ChatGPT enterprise usage data and produce a polished leadership ready rollout strategy for ChatGPT work. Something that you guys might actually do in your day to day. I want to identify usage patterns, high value use cases, and I wanna be able to prepare that all into, like, a document form. And so you can see now here, this is one of the big differences of how you would interact with work versus ChatGPT is is that we actually the model went in and it worked to the agent for about sixteen different minutes. It loaded different tools that were required, it read files, and it went down and actually identified both where the data that I wanted to reference that it believed to be the most accurate to produce this work. And it went in it also went through and identified what might those usage adoption trends be that would be interesting. And it ended up actually not only giving me the findings, so saying, you know, 91% are active, but only 47% use ChatGPT on the last few days. But it actually went ahead and created a document, that looks pretty good actually that I can open in our in app browser here. I didn't do you can see I didn't do anything to this at all. It was just a one shot document, created that actually went and summarized these findings. And so one of the core differences between Chat GPT and what you could do in work is in Chat GPT, it's not actually gonna go and create, within a chat this document or this artifact that could be used onward. Within ChatGPT Work, you it's gonna do that pretty much automatically. You can you can open it in anywhere you wanted to. You can export it to Google Docs if that's what you work in, or you can continue to leverage it through just a doc. And you can go and you can see the data that it's been, pulled out here. It's made a set of different graphs to visualize that data, and different operating implications. You could I could then open this up and edit it further. And so this just gives you a sense in terms of, like, the power just through, you know, one shot prompt, sixteen minutes of the agent working to be able to produce a deliverable that could hopefully start to accelerate, a task that you might need to do in your day to day. So that kinda gives you, like, a a sense of, like, what the capabilities are of our work. The other big thing for you guys to know is, you know, we we have first class place for plug ins, which is the agent gets more powerful as it can interact with your tools and data. And so we highly encourage you guys working through the security reviews of plug ins to enable them for your organization so that the agent can actually work and find the data where it resides, the templates that you like to use, and string that together to create meaningful work artifacts for you. Yeah. And in our experience working with customers, some of the first moments people will have with agents, be it work or Codex, tend to be connecting the agent to communication tools. Mhmm. The moment that you can ask your agent to help you draft documents or to help you summarize messages you've received or help with anything like that, things just, like, really light up for people. Mhmm. So I we always recommend starting there, and then you can, like, progressively add more plugins over time. Absolutely. One thing that I wanna show you before I get into the controls aspect just to illustrate what a control can look like to an end user is as I was preparing in our test workspace here as well is I actually wanted to go and see whether I could send an email to Julian who you guys might have interacted with to tell her that I'm ready to present for the admin webinar today. And so what you can say is that it actually worked for about, for fifty three seconds to try and find the connector here. And it actually told me that, the connected mailbox couldn't expose sending in this session. And so as part of this, I then said, like, no. No. No. I'm sure you can use the connector. Go ahead and do it. But it then followed up and mentioned that, you know, it didn't have the permission for the specific connector to be actually to do any send actions. It only had the permission for read only actions. And so this is something that's really important as you think about connecting these data sources is that you have the fine grain permissions for each of these plugins to specify what you want or not or not to happen within your organization. And so for this organization that we're working in today, we've specified that we don't want it to take any send actions. And as a result, I can even like, I'm trying to prompt around it. I'm saying, like, how do I reconnect this? I really wanna send this email, and it's coming back and telling me that it's not allowed. And if I did have the permissions, this is how I would be able to do it. And so it just kind of shows you, like, as an end user, the model is going to adhere to the policies that you set up and the permissions that you grant to these plug ins across your organization. That wasn't planned. That was just a tangent as I was preparing for it, but I thought that it'd be useful to share. So now why don't we go into, like, how do you think about controlling a rollout of this? So I'm gonna go ahead and and share different screen here, which looks like this screen probably where it looks really familiar to you guys. I imagine you spent a lot of time within the ChatGPT workspace settings today. Before we get into that, I do wanna also highlight, we spent a lot of time talking about the desktop app. This is also how chat and work materializes within the browser as well. So when you go to chatgpt.com and you sign into your organization, you'll now be able to switch between chat and work pretty easily up top here. And when you leverage work in the browser, it's still using the agenda capabilities and interacting with the plugins to be able to get work done. And so, again, most powerful is gonna be the desktop app, which we recommend, but also easiest to get started is also leveraging it within the browser. So going over from the browser to workspace settings, gonna first start about talking about like how do you provision access to this. So what you're looking at here is our permission and roles page. This is how we control turning on access for any of our features. And so you can first start within, you know, the workspace default role, but what we probably recommend is you starting with a, like, pilot, role that you wanted to be able to turn on access for. And there's two controls that you wanna pay attention to here first. So enabling work in ChatGPT mobile and web. So that was the toggle that we just reviewed on the browser. So you can this is an early access, so it's giving you guys the preview of it. As Alexander mentioned, it's off by default, but you have the ability to turn it on here. So you can decide to turn off or can turn it on. And when you do this, when you decide to turn it on, people within that custom role have access to it. There's also two other toggles that are dependent on that. It's whether or not you're gonna allow, Ork to use browser use and whether you're also gonna be provisioning network access associated with it. So you have a a fine grain control in terms of how the agent operates also within the web and mobile and browser. Can we talk more about why why would someone wanna say change allow or disallow network access in browsers? I can also talk about it. Yeah. Yeah. I mean, maybe probably, like, the the most interesting one is, like, whether you'd want it to be able to, like, interact outside of your network or whether you'd wanna be able to control it. But maybe if you wanna talk about, like, a more specific example. Yeah. So, you know, one of the one of the questions, with these kinds of agents is they're incredibly powerful. They can process a lot of information and do a lot of things for you. And we spend a lot of time training our models, to be really safe to use and secure to use. And, work mode has a powerful sandbox that you can you have a lot of configuration options over the sandbox. And, also, we have some different approval policies that you can set up. You can have your users approving actions that are taken, and, also, you can, you can have our we have a mode called approve for me where we actually have a model inspecting and making sure that what the model is doing is right. And so we ourselves work with network access and browser use enabled because these enable, the agent to be way more effective. Like, if, you know, imagine, as if as a teammate, if if I didn't have access to the Internet to do my work, I'd be pretty limited in terms of what I could do. You know, if I could only use, the certain connectors that we've enabled. That said, if you want to start, like, with the absolute most conservative policy, or there are certain kinds of data that you're, like, really concerned with, like, you want to, make sure that any agent that has seen that data is not gonna be able to access the Internet, then you can disable these toggles, and that would just get you to the most conservative, security posture. Our goal, like, as we build these products is basically to think about how can we make it as safe and as powerful as possible. And then we'll always share with you, like, this is how we work as OpenAI. We're we're pretty you know, obviously, we have the model weights. We're a pretty conservative company in this regard. And so we do a lot of work to to make the agent both powerful and safe. But at the same time, we always wanna give you an option. You can be more conservative or less conservative than us. Mhmm. So that's probably the first place to start is whether you wanna decide who has access. The other thing that I would highlight is that this turns it on for mobile and web, but we also have the capability down within the codex. Because as I mentioned that underneath the hood, whether you're using ChatGPT worker or you're using codex, it's still using the same, like, agenda capabilities. We have a single toggle here that decides whether or not you're enabling codecs and work locally within the desktop. And so, this is something that has been supported to date. And if you wanted to grant more people access to leveraging work within the desktop, you would use the toggles that exist here to be able to do the provisioning. And so, that's how I would recommend just getting started, while we're in this admin preview is being able to configure a role, give that role access to either use it on the desktop app or use it within web and mobile, decide whether you're comfortable using both the browser use and the network access to be able to provision that access. One other thing while we're here that should hopefully be like a big quality of life improvement as well if you haven't seen this already, we heard the feedback loud and clear that it was very difficult to know the who and what roles had access to what. And so you can now see the permissions across the roles and should be able to be pretty easily tell who has access or not, especially if you have lots of custom roles. So this is the very first place that I would start. I see a heart for that. I think you guys can expect a lot more changes and improvements to our back going forward. Yeah. Actually, we've been, we've been doing just, like, a bunch of great, like, UX work Mhmm. In the admin console. And something that we've been talking about is just like, hey. Should we, like, market this more? You know, we we don't think, all of you are on Twitter necessarily as much as, say, all our developers. And so we're we've been brainstorming, like, what is the right place to just share with you all these UI improvements that we're making. Yeah. So, also, feel free to drop, suggestions in chat. One one thing we're thinking about is, like, changing up the admin email Mhmm. To be a bit more visual and include some of this, but that might also make it harder to read. So we're curious what you think. Yeah. I just wanna say, like, you know, we shipped to ChatGPT 5.6. We shipped to work mode, but the thing that got the most reactions is the RBAC page. So, we hear loud and clear what you guys wanna see more of. And so this is, so stuff like this is what we'll So I mean, we so we literally have a term we use to describe this at OpenAI. We it's some extent we thought we felt like the capability overhang. Yeah. And and what we mean by this is, like, the models can be I mean, the models are incredibly smart. Right? We have the the leading, most intelligent, most efficient frontier models. And yet, right, for these models to be useful at work for all of your teams, actually, the enablement and the control of all of this is is probably the single biggest blocker. And so that's that's why we're so focused on it and why we're doing this webinar. Yeah. Absolutely. Cool. Okay. Other things that you guys probably care about is also on the plug in side. Mentioned, like, setting up and giving the agent access to your tools and data is a single biggest performance boost and the way in which you're gonna get the most leverage out of using these capabilities. And so what we have here is the ability for you to see all the different plugins that we support. These are we have, I think, over, like, a thousand or so plugins that are just in available to be installed. And then we also have the ability to create plugins, that you might have for internal tools, systems, data that you care about, where you have the ability to package it and share with specific roles across your organization. And so you can see, you know, like, a plugin like this, which is gonna be a custom plugin. It's enabled for the organization, and you can decide which specific, roles that you wanna be able to have this plug in installed for or available to. And so this is one of the biggest lists that we've heard where you wanted to be able to, centrally manage plugins that you create and share them with specific roles across your organization. But this interface is the way in which you would decide how to enable it and who should get access to it at the role level. This is also an area where we're making a lot more improvements to. So, you know, same way we shipped that table to show you better, controls in terms of who had access to what, we're making a lot of improvements to make sure that you guys can very easily decide who should have access to what plugins. And then also as part of the plugins, what are the specific permission scopes that you wanna be able to enable for it? One question I think we can answer just while we're here. A lot of the questions we'll answer at the end, but the difference between an app and a plugin. So just to explain quickly, plug ins for us are like a governable container for deploying capabilities to agents at your company. And so, you know, we when we were thinking about how to just, like, make things, both the most as powerful as possible for users, but also just as easy as possible to govern, We figured, okay. We just want, like, a single container that you and everyone can think about. And so, you know, you have this plug in interface. You can see all the plug ins at your company. Plug ins can contain a variety of different capabilities. So they can contain skills, which are instructions for the model, mostly instructions for the model to do things for the agent to do things. They can contain apps, which enable the agent to connect to other systems, such as Microsoft Teams. They contain they can contain hooks, which are basically a way for you to inject something that you want the agent to do under certain conditions during a rollout and more. So you have all of these different capabilities that you can use, but only the person who's making a plugin needs to know about it from everyone else, like all your customers and as an admin. All you have to think about is like, okay. What capabilities do I want shipped? And plugins are just the only thing that you have to, like, govern. Yeah. Exactly. So we're really, like, trying to consolidate that permission model where, like, plugins are the thing to care about. Alright. Other controls. I hope to get the same reactions, for this as I did for the RBAC table, but usage limits. One of the big things that we hear is we start to move from a world where, things are longer running agentic tasks. Okay. I see the hearts coming in. Yeah. So you're just finding too. This provides you the flexibility to be able to ensure that you're governing how spend is distributed across your distributed across your organization. And so just previewing what this looks like, we now have the ability to have a more flexible model than what we had before. Previously, it was weekly, tied to, RBAC. And what it is now is we have a dedicated page for usage limits, and we have a hierarchy of where you set a default monthly limit for the workspace. You then have the ability to set overrides at the group level, and then further, you can set overrides of the power or at the user level. And so the way to think about how this works in practice is you might want to enable for your organization a $100 100 credit limit, for example. And then for specific groups like engineers, you wanna be able to provide a higher limit on top of that. And then you might have power users within your engineering organization, and you wanna be able to further set an override so that those power users aren't capped. This gives you the flexibility of being able to do that entirely. The other feature that, you know, I think we haven't done as great of a job about probably talking about is that we also provided the ability to have, a customized the usage increase request process. And so what happens that for the end users, they'll get result they'll get, notification of when they're at 70% of their limits, 85% of their limits, and when they've hit their limit, and these are monthly limits. When they hit it, they can either just request, you know, someone within their admin to get additional usage, or you can use this functionality here to be able to decide how you want that request to be routed. So you can actually put in a custom description here and a custom URL. So if you have a policy to manage it in your SharePoint or your Confluence, you can actually update that here. So when a user hits their usage limit, they can go directly to the URL that you want them to go to, and then they can take actions subsequently to your policy. You can also have the ability to turn this off entirely. We You can also have the ability to turn this off entirely. We probably don't recommend that. Otherwise, users are just gonna be stuck. But you do have the ability to just customize what what it looks like. Dan is asking a question on this, which is could pending requests at least send an email when someone makes one or could? Yeah. Yeah. So, actually, that's something that we are planning on shipping tomorrow is the ability of when someone requests it through the platform. So if this is in default, we're gonna start to send batch notifications to the admin. So right now, they just show up kind of here in this pending request. But what you'll see is also a email showing you how many requests are pending and be able to go into the system to approve them. Other thing that is really big that we've heard your feedback loud and clear is all of this is gonna be accessible over API as well. We know for you guys that have, like, thousands of users across the organization, you don't wanna manually do this in the platform. You will be able to, like, set usage limits, be able to understand, change them at the group level. Like, anything that you can do in this UI, you will be able to do via API, and that is a stance that we're taking going forward. So any admin action should be available via API too. That was the third most. Okay. Yeah. We were keeping track of that. Yeah. So that's that's the second. The controls as we think about it, like first, figure out provision access, set up your plugins, make sure that you have your usage limit set so that spend is well within your control and that it's being governed. Then last is then Martin question on this last one. I'm just trying to the ones that are, like, topical I'm trying to do yeah. So Samus wants to confirm, did that does that mean emails are going to be sent to end users as well, and can we customize those both for when, you know, and also when, limits are approved or rejected? That's a good question. To start, what we're having is these requests are just going to the admin. So the user will be able to, like, have the information of where they are relative to their usage limit, and they'll receive, like, in product notifications when they approach their limits, so that they're getting it in context. But admins will receive the email. If there's a need for end users to receive it to you, like, we're very open. So just keep the feedback coming if we need that. Anything else? When does the API become available? Tomorrow. We're actually, as a sneak peek, tomorrow, we're launching, our ability to have spend control API, so everything that you can do here. Our cost or analytics API, you'll also have access to. And then the third is our group API. So if you manage any of your groups outside of SCIM, you'll be able to do that directly via an API. The other thing that you'll yeah. The other thing is that you'll be able to self-service create API queues. You know, that's been a big headache before as well where you had to go through support to do that. So, all of that's coming tomorrow, and this is just the first of many improvements. Do we want people to keep that low key or Not. No? Okay. Cool. Great. Okay. We're we're super excited about it. You guys should have it, and we should have built it already. So I think there's there's nothing low key about that. Okay. Next last thing that I want to show here is, analytics. So after you've provisioned usage, you've set your usage limits, you've connected plugins, you then want to be able to monitor how is your organization actually, leveraging work versus Codex versus chat. So What we are here is we're within the admin console. Think about this as like the if we're separating out the admin in the data plane, like console is the admin that cuts across all of OpenAI's products and will be eventually where we move all of the settings that you have within ChatGPT workspace into this single centralized place. So if you happen to be managing both your API org and your ChatGPT workspace, you have the ability to manage them into a single pane. So more we know you're in a fragmented state right now. We're in the process of consolidating that. So what you're seeing here is this, like, admin console. On this overview page, as part of this launch, we actually now have the ability to provide you, who's using chat, who's using codecs, and who's using work. And you can see the daily active users here. You can see how many credits. You can see a leaderboard of who are your top users and how many credits. If you click on any one of these, it'll actually be bring you down to a more filtered page for that given user. We're also working on performance for this as well so that it loads faster. But you can see this is how you get into, like, the individual user view here of how many credits they've used. So giving you, like, the fidelity of, you know, who's using what at multiple levels. So at your organization level and at your user level, how tokens are being spent across these different products and messages. And then we also give you more granularity if you wanted to go into the work and codex tab to see, you know, across, like, which models are being used, what surfaces. Also token breakdown by models, lines of code generated for anyone that's using codecs, ability to see what are your top plug ins. This is particularly interesting if you wanted to see, you know, what are the data sources that the agents connecting to most or that you should encourage other people to be using that might not be using and likewise with skills. So this is one of the the ways that we highly recommend you guys looking at and and getting, like, detailed fidelity of how this is being used. And then you also have the ability to filter down between worker codex so that you can see these same cuts. We're working on continually improving these views so that you have visibility into how, usage is trending across your organization. And also all of this will be available via an API as well is our goal. So we're gonna start with just the, like, credits data, tomorrow, but working increasingly on giving you these usage analytics via API as well. There's a bunch of questions in here related to this. Do you wanna just skim read this Yeah. On my computer or I don't know. And then you can just cherry pick a few to answer. Okay. Cool. How how far back data will go? So, tomorrow, we're sending the data to go back a hundred and twenty days for this overview. Right now, I think it's, like, thirty days, which has been pretty painful. So usually, you have access to about a hundred and twenty days worth of data. What is driving my credit usage? That's a good question. So if you want us to understand, like, how how is credit usage, being consumed, I think the way that I would actually probably think about leveraging that is looking at our dedicated credit page here. So you can see it by product, but then you can also see it at, like, high fidelity of the exact, like, model breakdown, so that you know exactly at the SKU level how kinda usage is being used. For admins, though, right, not for users? This is just yeah. This is just for admins right now. Yeah. But we are planning on bringing, more of this visibility to to users as well. Mhmm. I saw a bunch of requests for that. We don't have that for you tomorrow, but we are planning on that. Let's just see if there's anything else in here that you think. Dollar amounts, that's also something we're working on. I I know that, you know, you guys don't naturally think in credits, and so we're working on figuring out the estimation for dollars. That's a big ask. Alright. Let's keep going. I know Yeah. I think that that's actually Were you gonna show because there's a bunch of questions about restricting model use. Were you gonna show that, or do you want to? Yeah. So So the questions that I'm I'm summarizing here, I picked a number of people were asking how can I sort of decide which models, my team is enabled to use? So Hannah's just pulling that up. The answer is you can. Yeah. So, right now, we're we're working on improving our model controls pretty significantly. So, the first thing that we're working on supporting is actually, being able to have sticky defaults right now. And so What we're doing is within, Codecs and Work, what we've heard is that, you know, you might want, your users to use Fast Mode, but what happens is once they once they start using Fast Mode, every chat thereafter persists the user's defaults. One of the things that we're working on is being able to give you the control to say, what do you want each new thread or each new, thread or chat to start with? And so if you wanted it to start with a specific model, a specific reasoning, and a specific speed, you would then have the ability to do that so that your default persists versus the user defaults. And so this is actually in alpha right now. We're working on getting more feedback. If you're not in the alpha, it's a pretty large one, so reach out and we can get you gated into it. But what I'm gonna show Who should they reach out to? Whoever your account director is, and and they'll end up reaching out to us. So we'll we'll hear it. But, this is a our new dedicated model tab here. And so this is gonna be available within your, ChatGPT workspace settings. You can show it's in models. You didn't see it previously because it's in alpha, but on this page, you'll be able to set your defaults for codecs and work. with, the reasoning level, the speed, and then also whether the new chats this is the new behavior we're introducing, whether you want it the user's default to persist or if you want your admin default too. And I was expecting more emoji reactions to this. Come on. A lot of controls. A lot of controls. And then we're we're working on building this out, more as well. Thanks for the emojis. Alright. Okay. That was that was the last thing that I had. Okay. Let's go. Can we go to the slides? We I'll I'm going to blitz through. We had a few slides here on sort of rollout recommendations, but I'm gonna do this in about ten seconds, and let's mostly do q and a. Do you mind screen sharing it? Yeah. Okay. Awesome. I'll just do this Yep. Here. Okay. So some next steps we recommend. Honestly, the first and most important next step we recommend is trying it out yourself. We can talk as much as we want about what an agent is and how you use it and how it's more powerful than chat. But at the end of the day, the most powerful thing the the most sort of informative thing you can do is just check it out for yourself. So we'd recommend create a group and a role for yourself, turn on access for yourself and any other sort of immediate testers you want, and experiment with some low stakes things. For example, if you have a PDF with some SOPs on your computer? You could ask the the agent to answer a question about, like, hey. Like, what is our standing operating policy offer, suggest we do in this scenario? Or you could connect it to a tool like MS Teams and say, hey. Do I have any urgent messages to reply to? Once you've tried it yourself, the next step that we recommend is preparing a pilot. So here you go ahead and create a group. I'm gonna blitz over this. So, we we'll share the slides and the presentation so there'll be more there. But create a group, check out what plugins would you like to have access, folks to have access to and to connect to. Again, we recommend starting with communications tools. You're gonna wanna go check out usage limits. You have the controls that we just shared on usage limits, so you can really go in and customize. But set your usage limits and now send comms to your pilot group, and we're very happy to help you with this. Like, the way that we view our jobs of of, you know, ultimately making AGI beneficial to your teams is, first, on the product side, maybe that's 50%, build the right product, make sure it has the right controls. But, actually, the other half is really sort of the human side that you all have to engage with your teams on. And so we we have a lot of experience doing this, working with many customers on this. And if you reach out to your account director, we have a lot of recommendations for how to communicate with your team. And once you're running the pilot, collect feedback, share the feedback with us, please, with your AD, and then ultimately enable broadly. And that's it. So, just some quick resources we're going to send to you after this. We'll send you a bunch of resources for your customers, a checklist for a rollout guide and security, and just more information on setting up everything we showed you. And then also, we found a lot of these tools. You tend to have, like, one or two people on the team who are really passionate. We call these people champions, and so we have some resources for them too. And with that we have a couple minutes so let's do q and a. I think we've prioritized some. So this first question we're gonna take a few. Yeah. Yeah. The first question that is sorry this is a list that we can see not you all but was about how you restrict, model and thinking levels. We already answered that. So did we prioritize another one a number two question here? When will oh, great. I love Wow. When will we get so we're shipping, the So we're we're we're shipping the first version of the API tomorrow, which will be able to give you all of the what the user has used, what model reasoning, speed, and the associated credits associated with it. So that should be, like, some of the top asks that we've heard, but that should be coming tomorrow. Okay. I love it when the first few questions are things we're shipping. Okay. The last question. What is the cost difference using work and codex? Work and codex look quite similar. Generally, coding tasks are more token hungry than work tasks. So, yes, you should expect work to cost a little bit more than when you're using chat, but you have a lot of control over how that works. We will take a look at these questions, follow-up. There's a lot of questions in here, and I know we're over time, so we can't answer them, but we'll find a way to get answers to you all. So thank you so much for joining. We really appreciate this. That was awesome. We hope to do this more, so give us feedback. Yeah. Please let us know if we should do this more. Yeah. You know, we really don't know. Appreciate those of you who are much more actually just like, you know, we're here to help. We're excited to do this with you. So, yeah, just send us loads of feedback, and, yeah, we'll talk soon. Oh, thanks. I have