OpenAI launches ChatGPT agent

Two earlier strands

ChatGPT agent is best understood as the joining of two earlier lines of work. Operator was built to interact with websites, and deep research was built to gather and summarise information ^[1]. OpenAI described the new system as a unified product that combines Operator's ability to interact with websites, deep research's skill at synthesising information and ChatGPT's conversational fluency ^[3].

What launched

On 17 July 2025 OpenAI launched ChatGPT agent, which performs multi-step tasks for users ^[1][2]. The shift is from answering a single question to carrying out a sequence of steps — opening pages, looking things up, checking results, producing a file — that would otherwise be done by hand. An agent, in this sense, is a model given a goal and a workspace rather than a single prompt.

How it works

It ran on its own virtual computer. From there it could navigate websites, run code, conduct analysis and deliver editable slideshows and spreadsheets ^[3]. Like Operator, it controlled a virtual computer, and it inherited deep research's ability to gather and summarise information ^[1].

Users activated it by choosing “agent mode” from the tools dropdown ^[3]. While a task was running they could interrupt it or give additional instructions ^[1], take over the browser themselves, or stop the task at any point ^[3]. That supervision matters: a system that clicks and types on real websites can go wrong in ways a text answer cannot, and being able to step in mid-task keeps the work under the user's control.

Reported scores

OpenAI reported that the model behind it scored 41.6 pass@1 on Humanity's Last Exam, rising to 44.4 when up to eight parallel attempts were allowed ^[3]. It also scored 27.4% on FrontierMath with tool use ^[3]. The pass@1 measure counts a task as solved only if the first attempt is correct; allowing several attempts and taking the best one raises the number, which is why the two figures are reported separately.

Access

Pro users received 400 messages a month and other paid users received 40, with Enterprise and Education access promised in the coming weeks ^[3].

Why it matters

The launch marks a move from conversation to execution. A chat model answers; an agent is handed a goal and a machine to work on, and the interesting design questions move to supervision — when a person can interrupt, take over the browser, or stop a task ^[3]. By folding browsing, code, analysis and document output into one system that draws on Operator and deep research ^[1][3], OpenAI presented the agent not as a separate product but as a mode of ChatGPT itself, chosen from the tools dropdown ^[3].