OpenAI takes a new step with ChatGPT Atlas, a browser that natively integrates its artificial intelligence assistant.
Is this the new gateway to the web? In any case, this is what OpenAI wants. The San Francisco scale-up unveiled on Tuesday October 21 a web browser designed specifically to integrate ChatGPT. Like the main agentic browsers on the market, the AI can understand the data on the screen and act autonomously on behalf of the user. Better yet, Atlas remembers your browsing history to respond to your requests using data from sites you have visited in the past.
Currently the browser is only available on MacOs for free, Plus, Pro, Business, Enterprise and Edu users (with activation by the administrator). Everything is available worldwide, including in France (and this is notable). The JDN was able to test Atlas and gives you its first impressions.
A browser built around ChatGPT
The first big difference with a classic browser: ChatGPT is present everywhere. The OpenAI browser was developed around ChatGPT and it shows.
ChatGPT is the default search engine. It processes user queries, suggests relevant sites and now displays real SERPs (search page). Everything is very fast (faster than Perplexity on Comet, according to our tests) and displays truly relevant results.
As with the Perplexity assistant in Comet, OpenAI’s AI is controllable from a right-hand column window. The latter allows ChatGPT to be queried on the page currently being viewed. The AI has access to the entire web page but also to your complete browsing history (it is possible to deactivate this option). You thus have a real personalized assistant who knows you, your habits and your preferences.
Autonomous navigation: the real interest of Atlas
The real advantage of Atlas lies in its “Agent” mode. Comparable to that already offered in the classic version of ChatGPT, this mode is distinguished by the fact that the assistant now uses a local browser, and no longer a remote browser hosted in the cloud. This technical difference changes everything: actions are faster, more integrated and better contextualized on the user’s machine.
For example, it is possible to ask the browser to search for a product and add it to the cart on your behalf. The experience is quite smooth, certainly one of the best on the market. The agent really understands the context and achieves his goals without any problem.
Another use case, we asked the AI to sort our emails with Gmail: the AI was able to complete the task perfectly. A small message warns us, however, that the agent will only work while the page is open due to processing “sensitive data.”
OpenAI warns in its communication that the “Agent” mode in Atlas is an “early experience.” In other words, AI can make mistakes, especially during complex tasks. Finally, more importantly, there are real risks associated with using this mode. The agent may be the victim of a prompt injection hidden in an email or a web page in particular. It can then be diverted from its primary mission and leak data. Even with the hard work of OpenAI researchers to make it safe, agent mode can pose risks. It is recommended to use it, for some time, with human supervision.
What if the web was no longer the web?
At present, Atlas clearly outperforms Perplexity’s Comet, and this superiority can be explained above all by the total control that OpenAI exercises over its technological stack. The “Agent” mode developed by the San Francisco firm seems much more accomplished and native than that of Comet, which is explicitly based on a version of Chromium. Certainly, it is very likely that OpenAI also used Chromium as a base (the browser’s user agent testifies to this) but if this is the case, the code has obviously been partly, if not completely rewritten to offer a truly native experience. The immersion is undeniably stronger with Atlas than with Comet, and this is felt with each interaction.
However, Atlas will still need to mature before it can fully compete with traditional browsers. As it stands, several essential features are missing: for example, the browser does not have complete management of favorites, an integrated password manager, or even a file download manager.
Beyond raw performance, OpenAI’s ambitions seem quite clear: to become a new gateway to the web, or even to create a closed ecosystem in which it will be possible to do almost everything. Search for information, write, and even have fun with Sora, designed as an AI-generated video social network. The strategy is extremely intelligent: capture the user from their point of access to the web and keep them in an end-to-end controlled environment. A new stage, therefore, in the platformization of the Web and an additional nail in the coffin of the open web.
