OpenAI just launched “Operator”: Its first AI agent

• "Operator" can interface with websites to perform simple tasks

• OpenAI will release the model behind the agent via API soon

• The company plans to release more agents in the coming weeks

OpenAI has launched its first AI agent, "Operator": an assistant that can automatically perform everyday tasks like booking restaurants and filling out forms.

Operator "sees" and interacts with websites via a new "state-of-the-art" computer use agent model that runs underneath a dedicated cloud-based browser.

At the moment, it still makes "embarrassing" mistakes as it grapples with typing and clicking. But OpenAI expects the agent to improve fast.

CEO Sam Altman, who announced the release in a livestream Thursday morning, said his company had a "long and great history" of developing "early research previews" into "products people really love."

"This is really the beginning of this product," he said.

For now, Operator is only available as to Pro users in the U.S., but OpenAI will roll out out to Plus accounts in the coming months. Users based in the European Union may have to wait longer.

Under the hood

Operator combines a new "state-of-the-art" Computer Using Agent model with GPT-4o's vision capabilities. This means it can interpret graphical user interfaces like booking forms, and use "advanced reasoning" skills to fill them out.

It's not yet accessible via the OpenAI API, but the firm plans to release it soon so developers can use it in their own agents.

At the moment, users can ask the agent to perform tasks by inputting prompts into a browser interface, much like ChatGPT. The agent can then search the web and interact with the sites it finds to complete the job.

Although it's prone to errors, OpenAI says the CUA model is best-in-class. It gets web-based tasks right most of the time, according to the WebArena and WebVoyager benchmarks, on which it scores 58.1% and 87%, respectively.

CUA isn't quite so good at full computer tasks, which it performs successfully 38.1% of the time, according to the OSWorld test. But that's better than it's closest rival — Anthopic's Claude — which scored up to 22% on its launch in October. Claude does have an edge for developers, however, as it's available via an API.

Both Claude and Operator have been tested by a wide range of high profile companies to improve their real-world nouse. Anthropic has been used by Asana, Canva, Cognition, DoorDash, Replit, and The Browser Company before launch, while OpenAI is working with OpenTable, Etsy, Instacart, StubHub, Uber, Thumbtack, DoorDash, eBay, Target and others.

Safety first

OpenAI says it's rolling Operator out slowly because of how high the safety stakes are.

It's designed to always "ensure the person is in control" by asking questions and requiring confirmation before it performs tasks. The more sensitive the task, the more confirmation it will require. A prompt injection feature adds an extra layer of oversight designed to spot and stop suspicious activity.

But that doesn't mean users will have to wait for more exciting developments. Although it's been slower than other AI companies to release agents, the company says more are on the way in the coming weeks.