The most talked-about development in AI at the moment is AI agents, or systems that can perform intricate, multi-step actions on a user's behalf and go beyond chatbots.
OpenAI is fully embracing this trend by introducing ChatGPT Agent on Thursday, describing it as a tool that uses its own “virtual computer” to perform tasks on your behalf.
The new software is available to subscribers of OpenAI’s Pro, Plus, and Team plans. To enable it, users only need to select “agent mode” from ChatGPT's dropdown menu.
Here’s a breakdown of everything you need to know about ChatGPT‘s new software.
What is ChatGPT Agent?
ChatGPT Agent essentially functions like a personal assistant to ChatGPT, which you can delegate tasks to.
A monthly limit of 400 queries will be granted to ChatGPT Pro members whereas 40 enquiries will be sent to ChatGPT Team/Plus users each month.
Later this year, ChatGPT Enterprise and Education users will also gain access to the new software.
What can the new tool do?
According to the company, the new tool can do things like plan and buy items for a family breakfast, create a slide deck based on its analysis of rival businesses, and look at a user's schedule to notify them of impending client meetings.
The model that powers ChatGPT Agent, which goes by no name, was trained using reinforcement learning, the same method that powers all of OpenAI's reasoning models, on challenging tasks that call for a variety of tools, such as a text browser, visual browser, and terminal where users can import their own data.
According to OpenAI, ChatGPT Agent integrates the features of two of its current AI products, Operator and Deep Research.
Potential uses for ChatGPT Agent were showcased in a demo film shared on the Verge.
For example, to ask it to arrange a romantic night, it might connect to Google Calendar to determine the user's free evening and then utilise OpenTable to identify openings at specific restaurant categories.
The video also demonstrated how a user might halt the process by adding a different restaurant category to search for.
An additional example demonstrated how ChatGPT Agent could produce a study comparing the popularity of Labubus and Beanie Babies.
According to OpenAI, the model achieves cutting-edge results on a number of benchmarks.
This includes Humanity's Last Exam, where it received a score of 41.6 per cent, about twice as high as OpenAI's o3 and o4-mini.
Using a terminal that lets it run code, OpenAI claimed that ChatGPT Agent achieved 27.4 per cent on the most difficult maths tests.
In contrast, the o4-mini, which was regarded as FrontierMath's top scorer, managed to achieve 6.3 per cent.
