Software developer Tibor Blaho, known for accurate insights into upcoming AI products, has uncovered evidence of an OpenAI agent codenamed “Operator.”

Confirmed – the ChatGPT macOS desktop app has hidden options to define shortcuts for the desktop launcher to "Toggle Operator" and "Force Quit Operator" https://t.co/rSFobi4iPN pic.twitter.com/j19YSlexAS
— Tibor Blaho (@btibor91) January 19, 2025

According to Blaho, the macOS version of ChatGPT has hidden toggles enabling and disabling “Operator.” A user on X (formerly Twitter) nicknamed M1 independently confirmed similar details.

Blaho also found references to the AI agent on OpenAI’s website, along with a comparison to rival solutions.

Confirmed – the ChatGPT macOS desktop app has hidden options to define shortcuts for the desktop launcher to "Toggle Operator" and "Force Quit Operator" https://t.co/rSFobi4iPN pic.twitter.com/j19YSlexAS
— Tibor Blaho (@btibor91) January 19, 2025

The table suggests that “Operator” is not entirely reliable on certain tasks. In the OSWorld benchmark—an environment designed to mimic real computer tasks—the OpenAI agent scores 38.1%, outperforming the model from Anthropic but trailing a human’s 72.4%. Meanwhile, “Operator” exceeds human levels in WebVoyager, which gauges an AI’s ability to navigate web content.

The success rate for creating a Bitcoin wallet is 10%, while registering with a cloud provider is 60%.

Leaked charts indicate that the AI agent performs well in terms of security, as it resists attempts to engage in “illegal actions” or seek “sensitive personal data.”

Context