Meet OpenAI’s new web browsing agent.
On Thursday, OpenAI announced ChatGPT agent, a tool that’s capable of navigating the web and performing tasks on your behalf. As teased in an X post before the livestream, ChatGPT agent combines the autonomous capabilities of its Operator agent and the reasoning intelligence of its Deep Research tool.
OpenAI’s Operator, which launched in January as preview mode to ChatGPT Pro users, could scroll, click, and type on the web but had limitations and never saw a widespread release. Deep research is another type of agent that can search the web and compile information on the user’s behalf, but it couldn’t take actions beyond that. The launch of OpenAI’s new web browsing agent effectively combines both tools.
Credit: OpenAI
I’m testing Perplexity’s new AI browser Comet. Here are 3 features I tried right away.
“Operator couldn’t dive deep into analysis or write detailed reports, and deep research couldn’t interact with websites to refine results or access content requiring user authentication,” said the OpenAI announcement. “We saw that many queries users attempted with Operator were actually better suited for deep research, so we brought the best of both together.”
OpenAI’s new tool is part of the fast-growing agentic AI world
OpenAI’s ChatGPT agent joins other agentic tools recently released that can perform tasks on the user’s behalf. While not a full web browser, it acts similarly to Perplexity Comet’s browser assistant. Anthropic also has a tool called “computer use” that can take over your cursor and write code. As models become more advanced, they are more capable of performing autonomous tasks. Web browsing is considered one of the next arenas for AI labs to compete in, with OpenAI, Anthropic, Perplexity already shipping features, and Google’s Project Mariner research prototype.
Mashable Light Speed
When in agent mode, you can ask ChatGPT to perform tasks like “look at my calendar and brief me on upcoming client meetings based on recent news” or “plan and buy ingredients to make Japanese breakfast for four,” according to an OpenAI spokesperson.
OpenAI said that users are always in control and ChatGPT requests permission before taking actions, such as submitting forms, making purchases, or handling personal info. OpenAI said users can easily interrupt or take over when needed. Certain tasks, like sending emails, require oversight called Watch Mode.
On the safety front, ChatGPT agent is not allowed to perform “high-risk” tasks like financial transactions or legal advice, and it is trained to recognize malicious or adversarial attacks from prompt injections or other manipulative tactics.
In terms of data gathering, OpenAI said ChatGPT can delete their browsing data and log out of websites with one click and does not collect data while the user is actively involved in tasks like entering passwords.
Given the advancements, there’s also a considerably higher risk for things to go wrong. OpenAI has classified ChatGPT agent as high risk in its preparedness framework for biological and chemical capabilities.
There’s good news for users who don’t have the $200 a month ChatGPT Pro plan. It’s available to Pro, Plus, and Team users. Pro users get access today, and Plus and Team users get access over the next few days. Pro users get 400 messages a month, while Plus, Team, and other paid users get 40 messages a month.
Disclosure: Ziff Davis, Mashable’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.
