OpenAI Launches Atlas: A Browser Built Around ChatGPT

Just moments ago, OpenAI officially unveiled its first browser: ChatGPT Atlas.

For over a decade, Chrome has been the default gateway to the internet for most users. But what if there was an AI-powered browser with ChatGPT built in from the ground up, complete with memory and Agent capabilities? Would you give it a try?

That's exactly what Atlas aims to deliver.

A Browser Built Around ChatGPT, Not Just Embedding It

According to OpenAI, Atlas isn't simply ChatGPT stuffed into a browser—it's a browser fundamentally designed around ChatGPT. The core philosophy is comprehensive integration:

Sidebar Access Anytime: Click the "Ask ChatGPT" button to open the ChatGPT sidebar on any page. With full contextual understanding, ChatGPT can read the detailed content of your current page, help summarize information, explain concepts, or handle tasks directly within the browser window.

AI Writing Assistance: When typing in any text input field, ChatGPT provides real-time editing suggestions and intelligent auto-completion to boost your productivity.

Agent Mode: Enables ChatGPT to autonomously complete multi-step tasks within the browser.

Starting today, macOS users can access Atlas. It's available for ChatGPT Free, Plus, Pro, and Go subscribers to download and experience, with Windows, iOS, and Android versions coming soon.

Download link: chatgpt.com/atlas

However, Agent mode is currently only available to Plus, Pro, and Business users, and it's still in early stages—it may make mistakes when handling complex tasks. Additionally, if you set Atlas as your default browser, you'll receive a seven-day membership trial with higher usage limits.

Memory That Remembers Your Context

Atlas's memory feature is particularly interesting. It allows ChatGPT to remember the context of websites you've visited and recall that information when needed.

For example, you could ask: "Find all the job postings I looked at last week and summarize the industry trends so I can prepare for interviews."

This represents the complete vision of an "AI personal assistant"—one that remembers who you are, what you've done, and can complete tasks on your behalf.

Privacy Controls

When it comes to privacy, you have complete control over what ChatGPT can see and remember during your browsing. You can choose to clear records of specific pages, wipe your entire browsing history, or open an "incognito window" to browse temporarily without logging into ChatGPT.

OpenAI states that by default, your browsing content will not be used for model training.

Additionally, Atlas supports parental controls. If parents have set parental controls for ChatGPT, these settings automatically extend to Atlas. OpenAI has also added Atlas-specific parental control options, such as disabling browser memory or turning off Agent mode.

Atlas = Google + ChatGPT?

Opening Atlas's homepage reveals an interface quite similar to ChatGPT.

After entering a query, Atlas displays a selection window offering Google and Chat options, along with some specific links.

As the names suggest, the Google option redirects to a traditional search engine page, while the Chat option functions like a typical AI chat assistant, invoking the ChatGPT model to provide direct answers. Every question you ask in Atlas is saved to your ChatGPT account's chat history.

Beyond the answers in the dialogue box, you can switch tabs to browse search links, images, videos, and news, with most retaining the option to jump to Google search.

Performance varies between different models when asking questions.

For instance, when I asked whether pets can now ride high-speed trains in China, GPT-5 Instant failed to integrate the correct answer despite the official information being available in the search sources. In contrast, GPT-5 Thinking provided an accurate and rigorous response.

Interaction Design as a Highlight

Atlas's interaction design is a notable strength.

Traditional AI plugins typically appear as sidebars or floating windows, covering AI chat, translation, webpage summaries, and more.

Similarly, Atlas's search results page defaults to a split-screen display with ChatGPT. Clicking "Ask ChatGPT" in the upper right corner is like activating an AI plugin—the left side shows the original webpage while the right allows for multi-round conversations with ChatGPT, rather than traditional single-search redirects.

For example, highlight a word and the right side of the page can immediately search or explain it, or you can have it summarize an article.

During this morning's live demonstration, OpenAI employees searched for movies, viewed reviews from sites like Roger Ebert, then asked ChatGPT to summarize the reviews in five words or less—a task it handled easily.

It can even optimize webpage text and titles, with overall operations being quite smooth.

In today's demonstration, an OpenAI employee writing an email to a team designer selected a passage of text, clicked the ChatGPT button, and entered the "optimize language" command.

ChatGPT refined the text expression on the spot. Users can continue iterating modifications and finally click the "Update" button to replace the original content. This feature completely simplifies the previously tedious process of copying and pasting between documents and ChatGPT.

In my testing, summarizing video content worked without issues.

Atlas also supports screenshots. Click the + icon on the right, select "Attach Screenshot" to upload—I tried a photo of Elon Musk following these steps, and it accurately identified him.

Combined with the sidebar's Agent mode, the possibilities become quite interesting. A user let the Agent draw a flower on a painting website—the entire process took just 4 minutes.

It's worth noting that Atlas supports calling webpage tabs via @. However, unlike Dia which can call multiple webpage tabs simultaneously, Atlas can only call one open webpage at a time—a bit underwhelming, though likely to be updated later.

Agent Mode: The Main Event

Agent mode is Atlas browser's flagship feature.

In today's demonstration, an OpenAI employee planning to cook for 8 people opened a recipe webpage. He asked ChatGPT what ingredients were needed, and ChatGPT automatically analyzed the webpage content, calculated portions, and organized the shopping list by supermarket aisle based on the user's previously mentioned preferences.

After user confirmation, the Agent automatically opened the Instacart shopping website, searched for and added the required items to the cart—the entire process took just two minutes. While it didn't automatically place the order, having a prepared cart saved considerable time, allowing the user to review before deciding whether to purchase.

In another work scenario example, you can have ChatGPT open and read past team documents, conduct new competitive research, and compile it into a team briefing.

In practical testing, when I made a request like "Help me buy an iPhone 17 Pro Max on Apple's official website," it displayed "Take Control" and red "Stop" buttons while executing the task in Agent mode, allowing intervention at any time.

The execution speed isn't much different from the Agent mode in ChatGPT itself. While we might complete an order in a few minutes manually, having the Agent do it took 19 minutes—a speed that's clearly unsatisfactory.

In short, Atlas's Agent mode is essentially ChatGPT's Agent functionality extracted separately. Meanwhile, your chat history in Atlas is also preserved in ChatGPT.

Overall Experience: Pleasant Surprises, But Limited

My assessment after experiencing Atlas is: there are pleasant surprises, but not many.

Atlas feels like a fusion of Google search and ChatGPT conversation. However, Atlas retains ChatGPT's characteristic Agent mode, using ChatGPT's capabilities to re-integrate the browser scenario. Search, understanding, and task execution all complete within a closed loop.

It's worth mentioning that Ben Goodger, the engineer leading the Atlas project, has close ties to Chrome.

He joined Google in 2005 to participate in Chrome's development work and was one of the core members with the most outstanding contributions to Chrome. He joined OpenAI in 2024, becoming the chief technical lead for the ChatGPT Atlas browser.

Over the past period, various companies have successively launched their own AI browsers.

In this browser war, Perplexity Comet is more suitable for geeks or advanced users, while Atlas, based on the Chromium engine, has a lower barrier to entry and is more directly comparable to Dia.

Interestingly, Dia was recently acquired by enterprise software giant Atlassian for $610 million... so now it's Atlas versus Atlassian—quite a coincidental naming.

OpenAI's Product Strategy Evolution

Looking back at OpenAI's product philosophy, their previous strategy was to create a super app with all features packed in. But whether it's the recent standalone launch of the Sora app or now the Atlas browser, OpenAI is continuously spinning off functionalities from ChatGPT.

This aligns with Sam Altman's product thinking expressed in previous interviews: different products have different "mental positioning" in users' minds, and forcing them together destroys the experience.

Simply put, OpenAI's current strategy is clear: let ChatGPT play the role of a super operating system while allowing different products to serve their respective purposes, optimizing experiences for different scenarios. The transformation from "super app" to "product matrix"—OpenAI is determined to see this path through to the end.