OpenAI unveiled its latest model, GPT-5.5, on April 23, 2026, as part of its Safety Publication series. The model is designed for complex, real-world tasks, including coding, online research, document creation, and tool integration. Compared to previous versions, GPT-5.5 demonstrates earlier task comprehension, requires less user guidance, and uses tools more efficiently. It also includes self-checking mechanisms to ensure accuracy before finalizing outputs.
The company conducted full predeployment safety evaluations under its Preparedness Framework. This included targeted red-teaming for cybersecurity and biology risks, as well as feedback from nearly 200 early-access partners. OpenAI claims GPT-5.5 has the strongest safeguards to date, balancing misuse prevention with legitimate use cases. The model’s safety results are treated as proxies for GPT-5.5 Pro, which operates with parallel test time compute—though certain risks may differ due to this setting.
Evaluations were primarily conducted in an offline environment, except where noted. The System Card outlines these findings, along with details on the Bio Bug Bounty and cyber defense initiatives announced in April 2026. OpenAI emphasizes accelerating the cyber defense ecosystem while maintaining trusted access for future security needs.
The release follows OpenAI’s April 16 Security update on cyber defense access and the April 14 Security update on trusted infrastructure. The source URL for the GPT-5.5 System Card is provided by OpenAI.
Source: openai.com