TLDR : ChatGPT Agent can now act autonomously using its own virtual computer, handling complex tasks from start to finish while maintaining user control and security.
Table of contents
OpenAI introduces advanced agent capabilities in ChatGPT, marking a decisive turning point: the model is no longer limited to reasoning, it acts. By merging the features of Operator (active web navigation) and Deep Research (advanced synthesis), the ChatGPT agent can handle end-to-end tasks. It autonomously selects the most suitable tools, such as a visual browser, API, terminal, or code editor, and combines them into a single workflow, under the user's supervision.
Reason, Execute, Iterate
OpenAI states in its presentation:
"ChatGPT can now work for you using its own virtual computer, managing complex tasks from start to finish by smoothly transitioning between reasoning and action."
The agent operates in its own computing environment, allowing it to maintain context throughout the task. It can analyze competitors, generate presentations, fill out forms, or organize trips while remembering previous steps. The user remains in control of the process: explicit validation before each critical action, the possibility of interruption, the ability to review the action history, or make real-time adjustments. If a task requires an ID or password, for example, the ChatGPT agent pauses and asks the user to take control of the virtual browser, ensuring no screenshots are captured.
Expanded and Measured Use Cases
The agent targets both professional cases (automatic dashboard updates, financial modeling, assisted writing) and personal ones (reservations, planning, administrative assistance). Tests on several benchmarks confirm superior performance compared to previous models and even humans on some tasks, with a notable score of 89.9% on DSBench (data analysis) or 71.3% on SpreadsheetBench (editable Excel sheets).
Security and Governance
The deployment of these agent capabilities comes with safeguards. ChatGPT refuses certain sensitive requests, applies a strict hierarchy of instructions, and protects against adversarial attacks such as rapid injection. The agent operates transparently: users can see its actions and retain control of their data.
Availability
The ChatGPT Agent began rolling out yesterday for Pro, Plus, and Team users and will soon be available for Enterprise and Education users. The usage limits by plan type are as follows:
- Pro: 400 messages/month;
- Plus: 40 messages/month;
- Team: 30 credits/month.
Only user-triggered actions that change the agent's behavior count towards these limits: starting a task, interrupting a task, or responding to blocking questions. Users can increase their quota with flexible options based on credits.
Note that the ChatGPT Agent is currently not available in Switzerland or the European Economic Area (EEA) due to regulatory constraints.