OpenAI just combined two of its most interesting capabilities into a single agentic model, and the result is something that feels less like a chatbot and more like a junior analyst you’ve handed a laptop. The new ChatGPT agent merges the multi-step research chops of Deep Research — the feature that could spend 5 to 30 minutes generating a fully cited report — with Operator’s ability to actually use a browser, clicking buttons and filling forms on your behalf. That pairing alone would be notable. But the company didn’t stop there.
The agent also ships with a terminal tool for executing code, running data analysis, and generating output like slides or spreadsheets, though network access from that terminal is limited. First-party Connectors give it a direct line into external data sources and applications, starting with Google Drive. The throughline here is clear: OpenAI is building an agent that doesn’t just answer questions, but takes action across the tools knowledge workers actually use.
Safety, predictably, is front and center. The company expanded on controls established during Operator’s research preview and added new guardrails specifically targeting the risks introduced by broader user access and that terminal capability. OpenAI is treating this launch as High capability under its own Preparedness Framework for biological and chemical domains, activating the associated safeguards.
That classification deserves a beat. OpenAI explicitly states it does not have definitive evidence that this model could meaningfully help a novice create severe biological harm — which is the company’s own threshold for High capability. Yet it’s choosing to act as if the model has crossed that line anyway. A precautionary approach, as they’re framing it. It’s a rare admission of uncertainty in an industry that usually prefers its safety narratives to sound more certain than they actually are. Whether this sets a precedent for other frontier releases or turns out to be an isolated case of abundant caution will be worth watching.
