OpenAI’s Long-Awaited Agent Is Finally Here

OpenAI Official Introduction:

Operator is one of our first agents. These AIs can complete tasks for you independently—just give it a task, and it will execute.

Think of it this way: give Operator a shopping list, and it will autonomously buy everything on your behalf.

As you can see, the operator’s hands have left the keyboard; every action on the screen is completed by Operator itself.

It can even make restaurant reservations:

Right after Sam Altman’s livestream ended, OpenAI President Greg Brockman couldn’t wait to announce:

2025 is the Year of Agents.

This time, Operator went from announcement to availability immediately—though it is currently limited to Pro users. Yes, that’s the premium tier costing $200 a month (approximately 1,458 RMB).

After watching the livestream, netizens were thrilled, jokingly referring to it as “Crazy Thursday” (a Chinese internet meme implying a sudden windfall or treat).

But…

Operator is impressive, but it would be even better if it were open-source. DeepSeek and Meta, step up your game! (doge emoji).

Mastering the Browser Without Human Assistance

Talk is cheap; let’s look at Operator’s demo to see just how “independent” it really is.

It can operate on almost any website without human assistance.

For example, finding a recipe for clam linguine on Allrecipes and adding all the ingredients to my Instacart shopping cart?

Its logic for performing these actions mirrors human behavior: seeing images and clicking the appropriate buttons.

This differs from other agents that rely on APIs or programming interfaces; Operator reasons based on a text-based chain of thought.

Once the menu is confirmed, which store should be used for ordering groceries?

The human provides further instruction to use “Gus’s,” and Operator navigates to the corresponding website to place the order.

When encountering actions like login or payment, Operator hands control back to the user.

In user tests, one blogger discovered that if Operator is blocked by Reddit’s anti-bot measures, it will automatically add “Reddit” as a search keyword to find relevant posts.

Users can also add custom instructions for a personalized experience, such as setting a preferred airline when booking flights.

Operator allows users to save prompts for quick access on the homepage, which is ideal for repetitive tasks like restocking items on shopping websites.

It can run multiple tasks simultaneously, much like opening several browser tabs—for instance, ordering a personalized enamel mug on Etsy while reserving a campsite on Hipcamp.

At its core, Operator uses a new model called Computer-Using-Agent (CUA).

By combining GPT-4o’s visual capabilities with advanced reinforcement learning for reasoning, CUA can interact with Graphical User Interfaces (GUIs).

Operator can see the content of web interfaces and perform any action allowed by mouse and keyboard. This enables automatic operation without requiring custom API integrations.

If it encounters issues or errors, Operator can use its reasoning abilities to self-correct. If it gets stuck and needs help, it hands control back to the user.

CUA has achieved State-of-the-Art (SOTA) results on both the WebArena and WebVoyager benchmarks.

Currently, US-based Pro subscribers can use Operator at operator.chatgpt.com. Plus, Team, Enterprise users, and those in other regions will have to wait a bit longer, but OpenAI has promised to integrate these features into ChatGPT in the future.

OpenAI Enters “Level 3”

In July 2024, OpenAI released its “Five-Step Process from AI to AGI”:

• Level 1: Chatbots – AI can interact with humans via conversation.
• Level 2: Reasoners – AI solves problems at a human level.
• Level 3: Agents – AI can execute action-oriented tasks as systems.
• Level 4: Innovators – AI can develop innovative AI.
• Level 5: Organizations – AI can perform the work of an entire organization.

At that time, OpenAI stated it was still in Level 1 and approaching Level 2.

Now, with the release of Operator, Altman announced:

This is our beginning to enter Level 3.

Notably, as mentioned at the start, OpenAI subtly emphasized a key point: Operator is only one of the “first batch” of agents, not the sole agent.

During the livestream, Altman also teased:

We will be releasing more agents in the coming weeks and months.

One More Thing

Just before OpenAI’s livestream today, there was a small side story.

Two hours before the Operator announcement, OpenAI tweeted that it had fixed high error rates in ChatGPT and its API.

Another false alarm for netizens (doge emoji).

On a brighter note, Altman also teased that the free version of ChatGPT will soon have access to o3-mini.