OpenAI Just Released ChatGPT Agent, Its Most Powerful Agent Yet

OpenAI’s Isa Fulford, Casey Chu, and Edward Sun explain how merging Deep Research and Operator into a single RL-trained agent unlocks hour-long autonomous computer tasks.

• Agent runs tasks up to one hour on a virtual computer with text browser, GUI browser, terminal, and API integrations sharing a single file system state.
• Trained via reinforcement learning across hundreds of thousands of VMs; the model discovers when to use each tool without being explicitly programmed.
• Agent benchmark scores surpass human baseline on data science tasks; spreadsheet and slide generation are highlighted as new breakout capabilities.
• A demo task had the agent autonomously build an OpenAI valuation model with financial projections, a spreadsheet, and a slide deck in 28 minutes.
• Bio-risk was the top safety concern; weeks of red teaming were run, and a real-time monitor flags suspicious trajectories like an antivirus layer.
• Core team was 3–4 researchers on Deep Research side and 6–8 on computer-use agent side — combined for only a few months before shipping.
• RL is highly data-efficient: the fine-tuning dataset is orders of magnitude smaller than pre-training data, enabling fast capability iteration.
• Roadmap priorities: proactive agents that act without user prompting, personalization and memory, and new interaction paradigms beyond chat.

2025-07-22 · Watch on YouTube