OpenAI posts a system card for GPT-Live-1.1

The card is dated Oct. 15, 2026. The Deployment Safety hub lists “GPT-Live-1.1 system card” under updates for that day. GPT-Live-1.1 succeeds GPT-Live-1. Quip succeeds GPT-Live-1 mini. Full-duplex means the model can keep listening while it speaks, and can follow pauses and interruptions.

OpenAI says safeguards were built in largely the same way as for the earlier voice models: training the model to answer safely, measuring it on voice tests, and using the same monitoring and enforcement tools.

On synthetic voice tests, the company says GPT-Live-1.1 is close to its predecessor but has statistically significant regressions against GPT-Live-1 on sexual content, mental health, emotional reliance, and gore. It says more research is needed and that it plans to improve those categories. It says those tests do not include extra protections such as a trusted contact, crisis helplines, or parental controls. On a separate set of tests built from real user audio that people agreed to share, OpenAI says it saw no statistically significant regressions.

On its preparedness scale, OpenAI says neither model, when it is not handing work off to another model, could plausibly be considered High in biological and chemical risk, cybersecurity, or AI self-improvement. When a Live model does hand work to a stronger model, the company says the stronger model’s safeguards come along. It says the Live models themselves do not have broad tool access or the ability to run code at launch.

This story does not rank the model. The comparisons are what OpenAI printed.