Introducing ChatGPT agent: bridging research and action

Deep Research + Operator merged — neither alone was sufficient.
- Operator couldn’t write deep reports; Deep Research couldn’t act on sites.
- Same model family as o3; unification unlocks new capability class.
First ChatGPT product that takes real-world actions on the web.
Four tools: visual browser, text browser, terminal, direct API/connector access.
- Terminal has limited network access for code execution and data analysis.
- Connectors (Gmail, GitHub) let agent act inside authenticated sessions.
Prompt injection named the primary adversarial threat; model trained to resist.
- Malicious instructions hidden in webpages can hijack agent actions.
- Explicit user confirmation required before consequential actions (purchases, emails).
Rolled out July 17, 2025: Pro first, Plus and Team within days.