Google Chrome has a problem! OpenAI launched its own ChatGPT Atlas browser Home News OpenAI introduced ChatGPT Atlas – the first web browser with artificial intelligence directly at its core A key feature is a sidebar with ChatGPT that understands the context of the current page without copying text Agent mode currently only works for simple tasks, failing with more complex automations Sdílejte: Jakub Kárník Published: 22. 10. 2025 00:30 Advertisement On Tuesday, OpenAI introduced ChatGPT Atlas – a web browser with artificial intelligence built directly into its core. It’s not Chrome with an installed extension, but a new approach to how people use the internet, or at least that’s how the company presents the new product. While Google has dominated the browser market for over a decade, OpenAI is betting that AI will change the rules of the game enough for users to want to start fresh. Why a new browser at all? What Atlas can do – and what it can't Agent mode is not a miracle – yet Who has control over what The AI browser war has just begun When and for how much Why a new browser at all? OpenAI’s argument is simple: people today spend most of their time copying text from their browser to ChatGPT and back. You open an email, copy its content, switch the window to ChatGPT, paste the text, ask for an edit, copy the result, return to the email. The same applies to products on e-shops, recipes, work documents. Atlas aims to speed up this process by making ChatGPT available always and everywhere – in a sidebar that sees what you’re currently doing. When you open a recipe, ChatGPT can automatically compile a shopping list. When you read a work document, it can summarize it without you copying anything. When you browse job offers, it can create a summary of industry trends at the end of the week. Everything happens in the context of what you’re doing – not in an isolated chat application. What Atlas can do – and what it can’t Atlas launches with several key features. The most important is the sidebar with ChatGPT, which automatically opens when you click on a search link. You see the page and the chat simultaneously – you can ask questions, have text rephrased, or compare products without leaving the page. The second feature is browser memory. Atlas can remember which pages you visited and what you did on them, so it can offer more relevant answers later. This isn’t anonymous history – ChatGPT creates structured notes from your visits that it can use in future conversations. For example, if you searched for job offers last week, Atlas remembers this and can create a summary for you on demand. Memory is toggleable, visible in settings, and deletable along with browsing history. The third major feature is agent mode – ChatGPT can browse websites for you, click buttons, and fill out forms. Imagine you have a recipe and want to buy ingredients. You tell ChatGPT to open an e-shop, add items to the cart, and order delivery. Or you’re planning dinner and want to find a restaurant, book a table, and add the event to your calendar. In theory, it sounds great. In practice, it currently only works for simple tasks. Agent mode is not a miracle – yet OpenAI admits that agent mode is an „early experience” and that it can make mistakes with more complex workflows. Tests by TechCrunch and other media revealed that agents like Perplexity Comet or ChatGPT Agent handle basic tasks – adding items to a cart, finding information – but fail in more complex scenarios. When an agent needs to compare products based on specific parameters, fill out a form with conditional fields, or react to dynamic website behavior, it often doesn’t go well. OpenAI also warns of security risks. An agent can be attacked by malicious instructions hidden on a website or in an email, which overwrite its original behavior. This could result in data theft from logged-in sites or undesirable actions. OpenAI has conducted thousands of hours of testing but admits that protective mechanisms won’t stop every attack. Users should therefore consider what data they provide to the agent and ideally use it in logged-out mode on sensitive pages. The agent also cannot run code in the browser, download files, or install extensions. When working with banking and financial sites, it stops and asks for confirmation. These are preventive limitations designed to reduce risk, but they also limit usefulness. In its current form, the agent is not a universal assistant – it’s a prototype that is learning. Who has control over what Atlas gives users relatively detailed control over what ChatGPT sees. In the address bar, there’s a toggle switch that allows you to disable visibility for a specific page – ChatGPT won’t see it and won’t create a memory from it. You can delete your entire browsing history, which also deletes associated memories. Incognito mode temporarily logs out ChatGPT and disables tracking. OpenAI also states that by default, it does not use browsing content to train models. If you decide to allow it, you can enable “include web browsing” in the settings. This also applies to browser memories that inform your chats. If you have training enabled for chats in ChatGPT, it will also be enabled for Atlas. Parental controls from ChatGPT transfer to Atlas. Additionally, parents can disable browser memories and agent mode for children’s accounts. The AI browser war has just begun OpenAI is not the first to come up with the idea of an AI browser. Perplexity introduced the Comet browser this summer, offering similar features – an AI sidebar, video summaries, email inbox cleanup, or Amazon shopping. Google announced in September that in the coming months, it will integrate Gemini deeper into Chrome, allowing it to automate routine tasks such as grocery shopping, scheduling meetings, or booking restaurants. The Browser Company is working on the Dia browser, which has similar ambitions. The question is whether the world is ready to abandon Chrome with over 3 billion users for AI features that currently only work partially. In Silicon Valley, AI browsers are a hot topic, but in the real world, they are still a niche. Most people simply need a browser that works fast, reliably, and doesn’t interfere with their work. When and for how much ChatGPT Atlas is available from October 21, 2025, on macOS for all users – Free, Plus, Pro, and Go. Versions for Windows, iOS, and Android will come “soon,” with no specific date. Agent mode is available only for Plus, Pro, and Business users as part of a preview version. Enterprise and Edu users have access to Atlas if allowed by the plan administrator. Installation is simple – download Atlas from chatgpt.com/atlas, log in to ChatGPT, and import bookmarks, passwords, and history from your current browser. OpenAI promises frequent updates and publishes a roadmap that includes multi-profile support, improved developer tools, and SDK application integration. What do you think of the new Atlas browser from OpenAI? Source: OpenAI, The Verge, TechCrunch About the author Jakub Kárník Jakub is known for his endless curiosity and passion for the latest technologies. His love for mobile phones started with an iPhone 3G, but nowadays… More about the author Sdílejte: Atlas ChatGPT OpenAI Prohlížeč