INSANE ChatGPT 2025 Update — Agent Mode & o3-Pro Explained Fast

Imagine telling Chad GBT, "Hey, book my dentist appointment next week. " And it actually does that with agent mode. This is now possible. The agent mode is just one of many updates to Chad GBT we've got. And this video, I walk you through all of them. An autonomous coden buddy called, shiny new models, connectors, beefed up custom GBT options, upgraded projects, and so much more. So, how

about Chad GBT not only telling you what to do, but actually doing tasks for you online? That's what agent mode is all about. And it's probably the biggest update to Chad GBT ever. In simple terms, agent mode lets ChatGBT step out of just the chat box and take actions on your behalf on the web. It uses a sort of virtual computer with tools to browse websites, click buttons, fill forms, run code, even handle files, all while you supervise. It's not a new model, but a new capability built on top of the latest models. Of course, it's not like you'll kick back and let the AI run your life entirely. You stay in control. Agent mode gives you a realtime narration of what it's doing on screen and it will pause to ask for confirmation if something needs your approval. You can also interrupt it any time if it's going off track in any chat. Open the tools menu, the little drop down in the message box, and select agent mode. Boom. Chat GBT is now in agent mode for that conversation. Now you simply describe the task you want done. For example, you could say, "Find me three highly rated Italian restaurants in New York that have an open reservation this Friday at 7 p. m. and book one that fits my schedule. " Once you hit enter, the agent springs into action. It might search the web for restaurant listings, click into Open Table or a similar booking site, cross reference your calendar if you've connected one, and then actually reserve the table, all while showing you what it's doing in the sidebar. It's wild to watch. You guys know I'm all about leveling up your skills, and Skillshare has honestly been one of my favorite ways to do it. I recently went through their class called Chad GBT for Creatives: AI powered SEO, Marketing, and Productivity, and it's packed with seriously useful tactics. What I liked is that it doesn't waste time with fluff. In under 2 hours, you get realworld strategies on using Cad GBT to streamline your content workflow. stuff like brainstorming faster, writing SEO blog posts, optimizing video titles and thumbnails, even setting up AI to repurpose your long form content across different platforms. The instructor walks you through how to build a full content system using AI. And if you are a creator, freelancer, marketer, or just trying to be more efficient online, this one hits hard. So, if you're looking to make the most of AI tools, and I know you are since you're here, Skillshare has you covered. There are thousands of educational courses at Skillshare, marketing, video production, programming, anything you need. The first 500 people to use my link in the description or scan the QR code will receive a 1 month free trial of Skillshare. So, get started today. Big thanks to Skillshare for sponsoring this video. And all right, let's jump back into Chad GBT's new features because there's a lot to unpack. What can it do? Honestly, a ton. More than I expected at first. It can go out to the web, navigate pages, and read them either via text browser or visual browser. So, if you ask compare the specs of the top three noiseancelling headphones on Best Buy site, and give me a summary, the agent will actually click through Best Buy and scrape the info for you. The agent can also fill out forms online. It will always ask for confirmation before finalizing anything involving money or personal data, so you don't have to worry about accidental purchases. Find a dermatologist near me in New York that accepts new patients and book the earliest available appointment next week. The agent can search, find a booking portal, fill in your info, and get you booked. With connectors, it can read your email and draft responses or summaries. Summarize the important points from my last 10 work emails and draft polite replies. It will pull in your Gmail data if you've connected it. Then output summaries with cited snippets. Agent mode can check your calendar availability and schedule events. Schedule a team meeting for next Wednesday with Alice and Bob. If you've connected your Google calendar, it can suggest time slots and even create the calendar invite once you give the go-ahad. One of my favorite things is that it can generate and edit spreadsheets or slideshows for you. You can say, "Create a project timeline slide deck from the data in this spreadsheet, and it will produce a multi-slide presentation complete with charts or images. " Right now, the slideshow feature is still a bit basic in formatting, but it gets the structure right, and it's all editable in PowerPoint afterward. Perhaps most impressively, it plans dynamically. The agent can break a complex request into steps and decide which tools to use for each. If one approach fails, it can try another web browsing for info. Maybe using a connector to pull some internal sales data and making a slide deck. It can do that. Let's do a quick test. Say I prompt plan a oneweek trip to Tokyo for under $2,000 and prepare a dayby-day itinerary. Book flights and a hotel near Shinjuku and schedule any must-see attractions. The agent would start by searching for flights within my budget. Maybe use Kayak or Google flights via the browser. It finds some options, then pauses and asks me to pick one or it can pick the cheapest by default if I allow. After booking, it searches hotels in Shinjuku. Then it fills the booking form for a hotel and searches Tokyo attractions, compiles a list of popular ones, and maps them to the days of my trip, checking opening hours. Finally, it produces a neat itinerary document and even sets reminders on my calendar for each day's key event. Again, if I let it, all in one go with me just approving a couple of prompts. That's the power of agent mode. Now, before you rush off to try it, know that agent mode is not available for free tier. It's rolling out to pro plus and team plans. Also, it's currently georestricted in some places, not available in the EU and Switzerland. So if you're in those regions, you might not see the agent mode option. And because agent mode is pretty resource inensive, OpenAI sets monthly usage caps depending on your plan. If you are in pro, you get up to 400 agent messages per month. Plus in team users get 40 per month. And what counts as an agent message essentially every time you send a prompt that makes an agent do work or when you interrupt it or answer its clarification question. When you use agent mode, especially with your accounts, email, drive, etc., you'll need to connect those accounts via connectors and give Chad GBT permission to access them. The first time you'll go through a login flow for each service. And don't forget, you always remain in control. For instance, even though the agent can see your emails to summarize them, it's not going to start emailing people unless you explicitly instruct it. And usually, you'll have to take over for the final send. Data from connectors is used to read only by the agent, meaning can read and summarize info, but to actually act like sending an email or changing settings. It hands it off to you to confirm in the browser. But don't you think that prompting is gone now? No, now it's even more important. Nailing killer prompts is part art, part science, but you don't have to wing it alone. Our revenue focused prompt lab pro gives you 300 plus battle tested prompt and an entire track and automation productivity. Each prompt is paired with a note explaining the reasoning behind it so you can easily tweak it and apply it to agent mode or any other productivity tool. We use some of these prompt scenarios to saber marketing team 3 hours a day on email and you can do it too. Pick up the logic, tailor the wording to your own workflow and move faster. Whether you're polishing copy, fine-tuning outreach, or offloading repetitive tasks to an AI sidekick, agent mode shines for complex multi-step tasks, or anything that involves using external tools or websites. If your query can be answered with a simple paragraph from its built-in knowledge, you don't need an agent. Plain chat GBT is faster. But if you find yourself thinking, hm, I have the info. I just wish the AI could go and do XYZ with it online. That's the clue to unleash the agent. At the same time, remember its early days. The agent is powerful, but not perfect. It can sometimes get stuck or take a suboptimal route, but from my experience so far, it's surprisingly competent and only going to get better. You can always click the stop button if it's taken too long or going down a wrong path and you can wipe its browsing data and log it out of everything with one click in settings if needed. So don't be afraid to use it. Just keep an eye on it like a new intern who's very smart but might occasionally need guidance. If you've

been dreaming of an AI that not only writes code but also runs and tests it in your projects, Codeex is that dream come true. This isn't the old Codeex model. It's a new integrated coding agent inside ChatGBT. Think of it as a specialized AI developer that can plug into your codebase, make changes, and even propose commits or pull requests for you. Codeex is aimed at developers, engineers, or anyone working with code who wants to automate the boring or complex parts of software development. It can handle tasks like fixing bugs, adding small features, refactoring code, writing tests, updating documentation, answering questions about the code base, and more. The key difference from just chatting with GPT about code is that Codeex has direct access to your repository once you connect it and it runs each task in a cloud sandbox environment that mirrors your dev setup. This means it's not hallucinating what your code might look like. It's looking at the real files and can execute the code to verify its changes. In the Chad GBT interface, when you have a repository connected, you'll see codeex in the sidebar. You can assign a tasks by typing a prompt and clicking code for coding tasks or ask if you have a question about the code. Once connected, Codeex will download or index your repo into a workspace. You can then do things like scan the entire repository for any usage of a deprecated function and replace it with a new one. That could be a prompt or add a feature to export user data to CSV with tests. Codeex can open and edit files, compile or run the project, run test suites, etc. Really debugs and iterates on the code by itself until it's done. Each Codeex task is independent and runs to completion usually in one to 30 minutes depending on complexity. Once finished, Codeex commits the changes in its sandbox. It then shows you a summary of what it did along with citations of logs and test outputs as evidence. This transparency is awesome. It's like getting a pull request from a colleague where you can see exactly what they changed and their test results. If you're happy, you have a few options. You can directly open a GitHub pull request with those changes via button or you can download apply the patch to your local environment. Codeex is powered by a model called codeex1 which is essentially a version of OpenAI's 03 model optimized for software engineering. It has a huge context up to 192k tokens which is about 3,000 pages of code. So it's unlikely to forget parts of your codebase while working. You might be thinking, I already use GitHub Copilot or an AI in my ID like cursor. How is this different? Copilot and similar ID tools operate in real time as you code, suggesting lines or small refactors. They're like an AI parrogrammer in the moment. Codeex takes it a step further into the realm of autonomy. It's more like assigning a ticket to a junior developer. Hey, fix this book or implement this feature and let me know when you're done. You describe the task and Codeex figures out how to do it end to end. Running tests and verifying it works. If you have a GitHub repo, definitely give Codeex a spin. It's like having a super smart intern who can work 24/7 and loves writing tests and documentation. With all these new capabilities, you might

wonder what's powering them under the hood. Recently, OpenAI rolled out new models and some updates to existing ones. The new flagship model is the GPT 4. 1. Despite the name 4. 1 sounding like it might be a minor upgrade over the 4, it's actually a major improvement in certain areas. It excels especially at coding and long- form reasoning. The headline feature is a 1 million token context window. Yes, you heard that right, 1 million tokens. That's roughly 800,000 words. For comparison, the previous GPTs maxed out at 128,000. Beyond the context size, GPT 4. 1 has been fine-tuned for realworld tasks. It outperforms older GPT versions on many tasks. However, note that GPT 4. 1 does not have multimodal output, meaning it cannot generate images. It is focused on text and code. Alongside the GBT 4. 1, OpenAI also rolled out the 03 and 03 Pro model. The O3 Pro is easily OpenAI's most powerful reasoning model to date. It's extremely good at complex problem solving, code, math, science, and even visual tasks. This model is like the regular 03 on steroids and prompting works the same way here. Our generative AI course ships with a lean 03 prompting playbook battle tested through our own hands-on experiments. So you can slot it into your workflow and start seeing wins right away. The O3 Pro is basically the brainiest model when it comes to figuring out how to solve a task that isn't straightforward. For example, if you ask a complicated question like what are the latest developments in gene therapy for diabetes? Provide sources. 03 Pro might automatically use the web browser tool to find up-to-date info, then use its reasoning to compile an answer with citations all in one go without you specifically prompting it to browse. But because 03 Pro is so powerful, OpenAI did disable a couple features. It can't generate images and it doesn't support the canvas mode. Not big issues, just something to be aware of. You've got an arsenal of models. Now, I don't want you to get overwhelmed by the names. So, you better check out our video guide to all Chad GBT models we made just a few days ago. And don't forget, you can mix and match models in one conversation by just opening the model picker and switching models. I've already mentioned connectors several

times, and it's a relatively new feature, just a month old or something like that. probably didn't even notice them, but these are a huge addition that basically extends Cad GBT's reach into your personal and work data. Connectors are integrations that let Chad GBT securely connect to thirdparty applications and services like your Google Drive, Gmail, Dropbox, One Drive, Sharepoint, GitHub, Slack, and more. Think of them as bridges between Cad GBT and your cloud accounts. For example, if you connect your Google Drive, you could ask, "Search my drive for the budget spreadsheet from last August and summarize the key points. Chad GBT will actually search your Google Drive files for that query and pull the content to answer you. This transforms Chad GBT from just a general AI with public knowledge into a personalized assistant that can tap into your data and tools. Open AAI has two modes for using connectors in chat. The first type is chat search connectors. These are for quick on the-fly searches in a conversation. For instance, you can toggle on the drive connector and ask a question and ChadBt will fetch relevant files and info from your drive to answer. This is great for everyday tasks like what did that document say again or find me the slide deck about XY Z. Another type of connector is deep research connectors. This is a more heavyduty mode where Chad will use connectors to read through a lot of your data and produce detailed cited report. It's meant for more complex queries that might span multiple sources. Enabling connectors is pretty easy. Click on your profile icon in chat GBT and go to settings connectors. You'll see a list of available apps. Click connect next to one. A login window pops up for that service. Sign in and grant permission to open AI's app. It will clearly show what access you're giving. Once connected, it'll show up in your settings and you can disconnect anytime from the same menu. When you want to actually use a connector, you either start a new chat or open tools in an existing one. If you just want a quick search, click tools connectors. You'll get a list of sources. You can toggle on which ones to include for this query. If you want a comprehensive deep dive, click tools, deep research. Then you can select multiple sources and pick your connectors as data sources. Deep research might take a bit longer, but it gives you a nicely structured cited output, which is fantastic for reports or studying a topic. Always tell Chad GBT where to look. If you know the info is in a specific tool, mention it. Search or drive for the budget file or use the GitHub connector to find any PRs. You can also explicitly include or exclude sources in your prompt text at any time. You can manage connector data in the settings. There's a connection section and of course you can disconnect any connector anytime. Now, if you've

played with the custom GBTs feature where you can create your own chatbot personas, you'll know it was cool but somewhat limited in models. Well, not anymore. Custom GBTs now let you choose any of the available chat GBT models as the brain behind your custom bot. Custom GBTs are those little bots you can create and save each with a name, a custom profile, instructions, maybe some added knowledge or files, and a specific skill set. For example, you could create travel agent GBT that always helps plan vacations or code reviewer GBT specialize in giving feedback on code or something for fun like dungeon master GBT for RPG storytelling. Originally, all custom GBTs used the default GBT4L model that was fire. But sometimes you might want a lighter model. It doesn't need full power for a special reason in model. It's a logic heavy GBT. Now you have that control. When building or editing a custom GBT, you will see a model picker. Options include GBT40, GBT4, Mini and Nano, GBT 4. 5, and the O series models, 03, 04 Mini, and 03 Pro. If you're a pro, if you made a customer support GPT that answers common questions from a knowledge base, maybe GBT 4. 1 is overkill and a bit slow, you could choose 04 mini or GPT 4. 1 Mini for snappier responses, especially if you expect to handle lots of queries quickly. If you make a language tutor GBT that needs to explain grammar and check exercises, perhaps you want maximum correctness and multilingual knowledge, GBT 4. 1 full might be best even if it is slower and so on. Bottom line is custom GBTs are now much more flexible. I love this because it means I can tailor the AI to the task at hand both in personality and brain. If you haven't tried making a custom GBT, give it a shot. It's like creating your own mini chat GBT that you can even share with others and now you can ensure it runs on the best fit model. Now let's talk about projects. Another feature

that just got a major upgrade. If you haven't used projects in Chat GBT yet, think of a project as a workspace where you can keep a bunch of related chats, also files and notes all together for a long-term task. It's incredibly useful if you're working on say a research paper, business plan, learn a topic or any multi-step project where you might have different threads of conversation, reference documents, etc. And you want Chad GBT to keep track of the context across sessions. As you see, we have a few projects already. Each one fits his goal, whether it's script writing, creating images for thumbnails, or something else. We've spent tons of time perfecting our pipeline just to have each project solve a specific task. All those battle tested playbooks now live inside the AI master membership. Crack open the AI 101 crash course and you'll get bite-sized video walkthroughs, prompt packs, PDF cheat sheets that take you from doodle to polished asset minus the trial and error. Everything's modular, always current, and snaps right into whatever tool stack you're rocking. The payoff is better texts, cleaner images, and roughly a 30% drop in turnaround because the AI finally pulls its way. And since you're already here in AI master, you can snag 63% off your first year. Hit the link below, dive in, and let's turn those big ideas into production ready visuals today. You can now use the deep research mode within a project chat. This is awesome because it means the AI can research things with awareness of your project context. For example, if your project is about market analysis for electric vehicles, you might have notes and chats already. Now, you can say, "Run deep research on the latest EV sales figures and compared to my notes, and ChatGBT will use the web and your connected data knowing what's already in your project to give a comprehensive answer. And you can now talk to Chad GBT in your projects. You'll see a microphone icon in project chats. Now, seems easy. I'm sure you know how voice mode works. We had a whole ChatGBT guide where we covered that. Be sure to check it out. Within a project, ChatGBT can now reference past chats in that project as context. Essentially, each project has a memory silo. It will remember relevant details from earlier conversations in the same project. Even if you start a new uh chat thread in that project, essentially projects are evolving into your personal AI war room. You have all your data, persistent memory, voice conversation, and research tools all in one place. Makes Cad GBT far more useful for serious work. No more splitting info between multiple chats and losing track. Projects keep it organized and remembers context over days or weeks. If you haven't tried projects, here's how to create one. On the left sidebar, click new project. Give it a name. Inside, you can add a description or custom instructions if you want. Then you can start a chat within the project. Works like normal chat but with that extra context. Click project files to upload files to the project and instructions to edit how Chad GBT behaves in this project. Chad GBT is no longer just a

chat toy that forgets everything. It's becoming a real platform, an assistant that can act, remember, and integrate with our lives. Whether you are an AI beginner just excited to have ChatGBT book your flights or an enthusiast ready to automate coding tasks and set up custom AI bots, there is something here for you. If you found this video helpful, give it a thumbs up and consider subscribing for more AI tips and deep dives. Thanks so much for watching. Heavy experimenting with the new Chad GBT and I'll catch you in the next video.