What happened
OpenAI published (2026-10-15) the deployment-safety system card for its GPT-Live-1.1 and GPT-Live-1.1 Quip voice models, the successors to GPT-Live-1 now default for Plus-and-above users including in Codex and Dots. The card discloses that tests on real-world audio prompts found no statistically significant regressions versus predecessors, but synthetic evaluations flagged statistically significant regressions in several categories — sexual content, mental health, emotional reliance, and gore — which OpenAI says need further research. It also details the voice-specific safeguards, monitoring, review and enforcement infrastructure carried over from prior models.
Why it matters
The synthetic-evaluation regressions in safety-relevant categories (notably mental health) are a data point for organisations deploying voice assistants in consumer-facing or sensitive settings, illustrating that UI safety testing can diverge from stress-case evaluation results.
Action needed
Track whether OpenAI's follow-up research addresses the flagged synthetic-evaluation regressions before relying on GPT-Live-1.1 family models in mental-health-adjacent or unsupervised voice use cases.