OpenAI showed ChatGPT agent — it's Deep Research and Operator in one bottle

The new agent can multimodally browse web pages, call APIs and tools, and perform tasks with reasoning.

Special emphasis on using various tools — the agent was specially trained through RL to work with tools. It creates diagrams, presentations, generates images, can log in to sites and use the terminal.

The result on Humanity's Last Exam is 42%, which is a serious jump compared to o3 and even Deep Research. There is also noticeable progress on Frontier Math.

It's cool to watch such breakthroughs