OpenAI GPT-5.5 System Card

OpenAI has released GPT-5.5, a model designed for complex, real-world work such as coding, online research, and cross-tool orchestration. This release marks a shift toward models that require less guidance, use tools more effectively, and autonomously check their own work to complete tasks.

Model Capabilities and Task Execution

GPT-5.5 is engineered for high-autonomy task completion. Compared to previous iterations, the model demonstrates improved ability to understand tasks early in the process and requires less user guidance to reach a goal. Key capabilities include:

• Complex Workflow Orchestration: The model can write code, conduct online research, analyze information, and generate documents and spreadsheets.
• Tool Integration: GPT-5.5 uses tools more effectively to move across different applications to achieve a goal.
• Self-Correction: The model is designed to check its own work and persist until a task is completed.

Safety Framework and Predeployment Evaluation

OpenAI subjected GPT-5.5 to a comprehensive safety suite, including the Preparedness Framework and full predeployment safety evaluations. To mitigate risks associated with advanced capabilities, the model underwent targeted red-teaming specifically for cybersecurity and biology.

Before the official release, OpenAI collected feedback from nearly 200 early-access partners to evaluate the model's performance in real-world use cases.

GPT-5.5 vs. GPT-5.5 Pro

OpenAI distinguishes between GPT-5.5 and GPT-5.5 Pro. GPT-5.5 Pro uses the same underlying model but employs a setting that utilizes parallel test-time compute. Because this specific setting may materially impact risks or the necessary safeguard posture, OpenAI evaluates GPT-5.5 Pro separately in certain cases.

Generally, the safety results of GPT-5.5 serve as strong proxies for the safety profile of GPT-5.5 Pro.

Deployment Safeguards

OpenAI is deploying GPT-5.5 with its strongest set of safeguards to date. These measures are designed to reduce the potential for misuse while ensuring that the advanced capabilities of the model remain available for legitimate and beneficial uses.