Official ChatGPT Agent System Card
OpenAI’s ChatGPT Agent System Card, published July 17, 2025, is also available as a PDF. It describes the ChatGPT Agent launch: a product combining multi-step research, a visual browser, a terminal with limited network access, and connectors to external data sources and applications. These are descriptions of that ChatGPT product launch, not a guarantee that an API offers the same experience. OpenAI’s launch announcement says the launch was treated as High capability in the biological and chemical domain under its Preparedness Framework, with associated safeguards activated.
What the card reports
The card’s listed coverage includes standard safety evaluations; product-specific risks such as prompt injection, agent mistakes, and harmful or disallowed requests; red teaming; and Preparedness Framework capability assessments. Its listed product-specific mitigations include safety training, automated monitors and filters, confirmations in sensitive visual-browser contexts, terminal network restrictions, disabled ChatGPT memory, and watch mode. These are launch-product safeguards; their presence does not establish that every workflow is error-free or that an API application inherits them. See the card’s mitigation sections.
For factual question answering, the card reports that ChatGPT Agent scored lower than o3 on SimpleQA accuracy. It qualifies the comparison: the table compares Agent with o3 using browsing, while the o3 system card’s figures reflected performance without browsing. The card also notes that manual review surfaced potential grading-rubric flaws in some cases, including possible inaccuracies in Wikipedia content. This is a task-specific result with an evaluation caveat—not a general accuracy score across agent workflows. The launch card gives the comparison and qualification.
A separate July 2025 commentary by Zach Stein-Perlman discusses the card’s capability and safeguard claims. Treat its critiques as the author’s interpretation, not as OpenAI’s conclusions; use the official card for OpenAI’s stated findings.
Does ChatGPT Agent map to a production API?
Do not treat the former ChatGPT Agent product as an API model or assume its product safeguards carry over. The documentation summarized here does not establish a one-to-one API replacement for the ChatGPT experience.
OpenAI’s API model documentation identifies GPT-5.4 (gpt-5.4) as supporting computer use and lists the Responses endpoint. That documents a model-and-tool option to evaluate; it does not establish equivalence with ChatGPT Agent. Check the GPT-5.4 API page and current model catalog for current availability and model details.
The computer-use integration guide describes a developer-managed tool loop, including maintaining the environment between calls and returning screenshots. The available documentation reviewed for this summary does not establish all consequential access or usage restrictions, or whether those conditions make the integration suitable for a particular production deployment. Verify those conditions in the current guide and assess the environment and safeguards for your application before selecting it.
Bottom line: Use the system card to understand OpenAI’s 2025 ChatGPT Agent launch, its evaluations, and its stated safeguards. For API-based computer use, evaluate the documented API model and integration separately; do not infer that the ChatGPT product’s capabilities or protections transfer to your application.
Sources
ChatGPT Agent System Card — OpenAI Deployment Safety Hub
ChatGPT Agent System Card — PDF
ChatGPT Agent System Card — launch PDF
ChatGPT agent System Card — OpenAI