OpenAI today introduced Operator, an artificial intelligence agent that can automatically perform tasks on users’ behalf.

Two of the company’s highest-profile rivals announced their own product updates in conjunction. Perplexity AI Inc., a startup with a popular AI search engine, introduced an agent similar to Operator for its Android app. Anthropic PBC, which already offers such automation capabilities, debuted a tool that will enable its AI models to include better citations in prompt responses.

OpenAI’s new Operator agent is initially available in the top-end Pro tier of ChatGPT as a research preview. It can order groceries, book flights, fill forms and perform other multistep tasks. Users can instruct Operator what tasks to perform by entering natural language prompts.

Under the hood, the agent is powered by a newly revealed OpenAI model known as CUA. It’s partly based on the company’s multimodal GPT-4o large language model. OpenAI says that CUA combines the LLM with “advanced reasoning through reinforcement learning.”

When users ask Operator to perform a task in a website, the agent navigates to the relevant URL using a built-in browser. It can type, click and scroll to carry out the requested action. Operator regularly takes screenshots of the website to check that everything is working as expected. 

The user can take over at any point during the workflow, OpenAI detailed. Operator proactively asks users to switch to manual mode for sensitive actions such as entering login credentials into a webpage. According to OpenAI, the agent stops taking screenshots until the task is completed.

The company has built several data protection features into Operator. Users can log it out of all their accounts with one click and prevent OpenAI from using their data for AI training. Additionally, there’s a system that detects when malicious websites attempt to trick Operator into disclosing sensitive data.

Some of the agent’s features are customizable. A user could, for example, save a shopping list and have Operator buy the specified items every time it visits a certain e-commerce site. It’s also possible to create customization settings that apply to all the websites the agent visits. 

Going forward, OpenAI plans to expand the availability of Operator beyond ChatGPT Pro to the chatbot’s other tiers. The company will also offer the agent through its application programming interface. Under the hood, OpenAI plans to add enhancements that will make Operator better at completing complex tasks.

“Operator is currently in an early research preview, and while it’s already capable of handling a wide range of tasks, it’s still learning, evolving and may make mistakes,” OpenAI researchers wrote in a blog post. “Early user feedback will play a vital role in enhancing its accuracy, reliability, and safety.”

OpenAI rival Perplexity AI today debuted an agent of its own, Perplexity Assistant, that is accessible in its Android app. It can make e-commerce purchases, book a taxi and perform other tasks in an automated manner. A multimodal processing feature enables Perplexity Assistant to analyze smartphone camera footage and the content on the user’s screen.

On launch, the agent can perform actions in Spotify, YouTube and Uber along with email, messaging and clock apps. Perplexity AI plans to add support for more services over time.

Anthropic, another OpenAI rival, also announced a product update today. The company provides an enterprise-focused LLM series called Claude through an API. Using a newly added feature called Citations, customers can now upload documents to a Claude model and have it highlight the specific sentences it uses to generate prompt responses.

Image: OpenAI