OpenAI launches Operator — its first AI agent

 

After much talk, OpenAI finally launched an agent — an AI system that can do work for you on its own. It’s called Operator and uses a remote web browser in the cloud to do (or more accurately, tries to do) the tasks that you give it.

Under the hood, Operator is using a specialised AI model, Computer-Using Agent (CUA), based on GPT-4o.

I’ve gone down a bit of a rabbit hole on X to find the most impressive things people are doing with Operator:

Doing a bunch of daily tasks: ordering pizza, buying groceries, etc. (OpenAI’s own launch video). Building and deploying a tool on Github using Replit Building stuff on the no-code code app builder Bubble Creating a Google Form survey and testing it Putting together a folder structure for an M&A Solving a CAPTCHA Crypto research (when faced with a CAPTCHA to verify “Are you human?”, it pings the user to complete it!) Have multiple instances of Operator do work at the same time Planning a weekend trip based on hidden gems from Reddit What happens when you try to make Operator go on the web and use another instance of Operator (jump to 06:15)

The coolest thing about most of these demos isn’t necessarily the end result, but observing the way Operator actually works, especially when it faces obstacles (like being banned from certain websites and finding creative workarounds).

I know you’re all eager to test Operator out—but unfortunately for most— they’ve only made it available on ChatGPT Pro in the US to start with (yes, that’s the $200/month subscription). However, the plan is to expand access to more countries as well as ChatGPT Plus soon; Sam Altman also committed to launching their next agent with availability on Plus tier.

Operator is far from perfect, but it’s a big step forward for AI and one that takes it beyond mere knowledge—it now understands pixels on the screen and can use a virtual keyboard and a mouse to do things with that understanding.

Right now, the use cases flourishing on the web are centred around small, daily tasks (reserving a table, booking a flight, buying groceries). These can definitely provide small efficiency gains for daily life, but make no mistake. This is an intelligent agent that can do WORK for humans on the WEB. I think at the end of this year we’ll have a lot of people automating repetitive, boring tasks they previously had to do as part of their job with this and similar tools.

Essentially, every website is now available to capture information from, even though it doesn’t have an API. And Operator-like agents could soon start writing their own API for repeating tasks.

A significant obstacle these agents face, though, is that they’re (understandably) getting banned by website owners e.g. YouTube and Reddit. It’ll be interesting to see if the AI companies will find ways to bypass these security measures or have to start paying for access.

2025 was predicted to be “the year of AI agents”, and with Operator now out in the wild, that statement seems pretty likely. The other leading AI companies are going to follow suit.