GPT-5.5 System Card: OpenAI publishes safety evaluations and risk assessment for the new model

Alongside the launch of GPT-5.5, OpenAI published an accompanying System Card on April 23, 2026 — a technical document describing capability evaluations and safety measures taken before the public deployment of the model.

System Cards have become an industry standard since OpenAI introduced them with GPT-4 in 2023. Similar practices have been adopted by Anthropic (with so-called “model cards” and Responsible Scaling Policy reports), Google DeepMind, and xAI. For GPT-5.5, the document was published simultaneously with the model.

What is typically found in a System Card?

A System Card typically covers multiple evaluation areas. Capability evaluations measure the model’s performance on benchmark tests — from general reasoning to specialized domains such as mathematics, science, and programming. Bias and harm testing assesses the model’s tendency toward harmful responses, stereotypes, or misinformation.

A separate section is usually dedicated to red-teaming — controlled attempts by external researchers to “break” the model by requesting dangerous information or bypassing safety measures. For GPT-5-level models and newer, this typically includes evaluations of cyber capabilities (can the model help write malicious code), persuasion risks (can it manipulate people in sensitive contexts), and biological and chemical hazards.

Why is the System Card especially important for GPT-5.5?

GPT-5.5 is one of the rare models that comes with its own Bio Bug Bounty program — a public invitation for red-teamers to find “universal jailbreaks” in the field of biosecurity. This is a strong signal that OpenAI’s internal classification places the model in the elevated-risk category for dual-use scenarios.

In that context, the System Card serves as a reference point for regulators and customers: it demonstrates that controlled evaluations were conducted before deployment, that risks were quantified, and that mitigation measures exist. For compliance teams in regulated industries, the System Card is often a prerequisite for assessing model adoption.

What can researchers and developers expect?

For academic researchers, the System Card is the closest source available to them without access to OpenAI’s internal evaluations. Based on it, independent replications are typically written, comparisons with other models made, and safety guardrail analyses conducted.

For developers integrating GPT-5.5 into their products, the document helps in the risk assessment process — it identifies domains where the model needs additional guardrail mechanisms (content filters, rate limits, human-in-the-loop checks). This is especially important for startups building vertical AI solutions in healthcare, law, and finance.

The detailed numerical results from the GPT-5.5 System Card will likely be the subject of analyses in the coming weeks, as was the case with previous OpenAI models.