GPT-Live-1.1 reaches ChatGPT Voice
On 15 October 2026 OpenAI began rolling out GPT-Live-1.1 and its smaller GPT-Live-1.1 Quip in ChatGPT Voice, with more natural accents, a wider range of short phrases while it searches, and new voices in French, German, Hindi, Japanese, Korean and Mexican Spanish plus an English voice, Bubbie. The same day OpenAI published a system card in which its own tests found statistically significant regressions against GPT-Live-1 on synthetic prompts about sexual content, mental health, emotional reliance and gore.
Why it matters
System cards usually report parity or gains; this one reports losses, says more research is needed to understand them, and says that OpenAI plans to improve in these categories in future releases. The losses appear only on hard synthetic audio prompts, while the test on production-style samples shows no significant regression, and the card notes that its evaluations leave out system-level protections such as Trusted Contact, localised crisis helplines and Parental Controls.
What the card says. GPT-Live-1.1 and GPT-Live-1.1 Quip are the successors of GPT-Live-1 and GPT-Live-1 mini: voice models that are "full-duplex", meaning they can listen and respond continuously, following pauses, interruptions and changes of pace. GPT-Live-1.1 is the default voice model for Plus and higher tiers, including for Codex and Dots; Quip is the default for others. The card says GPT-Live is "actually a two-model system" and that, by OpenAI's hypothesis, its alignment behaviour is inherited from the "thinking" models behind it. Safety evaluations. On production prompts (real voice samples from users who chose to share them; not prevalence-weighted and built around cases where earlier voice models were not yet ideal) in six categories the models "perform at parity with their predecessors, with no statistically significant regressions". On synthetic audio prompts in eight categories the card says the models perform "close to parity" but that GPT-Live-1.1 "has statistically significant regressions on sexual, mental health, emotional reliance, and gore categories compared to GPT-Live-1": sexual 0.87 to 0.84, mental health 0.70 to 0.63, emotional reliance 0.86 to 0.81, gore 0.76 to 0.69; other categories barely moved (illicit behaviour 0.94 to 0.93, personal data 0.95 to 0.95, self-harm 0.89 to 0.87, hate 0.96 to 0.96). For Quip against GPT-Live-1 mini the table gives 0.66 to 0.62 on gore and 0.60 to 0.59 on mental health; the text of the card names only GPT-Live-1.1. The card writes: "more research is needed to understand the regressions, and we plan to improve in these categories in future releases"; these evaluations "do not account for system-level interventions" such as Trusted Contact, localized crisis helplines and Parental Controls for younger users. Alignment. On a test of respecting warnings the models have "slight regressions" against their predecessors; on blocker deception GPT-Live-1.1 does slightly better and Quip slightly worse; on a broken search tool GPT-Live-1.1 does slightly worse and Quip slightly better, and all GPT-Live models improve on GPT-5.6 Terra, the underlying model of these evaluations, "but more research is needed to understand the cause for that improvement". Preparedness. Neither model, operating without delegation, "could plausibly be considered High" in biological and chemical risk, cybersecurity or AI self-improvement; the voice models have no code execution and no broad independent access to tools, and OpenAI says it will reassess the cybersecurity safeguards before enabling additional tools. What the record does not claim. All evaluations are OpenAI's own, on OpenAI's own graders; the scores of earlier models are from the latest versions of those models; the numbers do not describe rates in real use, because the evaluations are deliberately hard and not prevalence-weighted. Quip's values in the four named categories are in the table, but the card's text names only GPT-Live-1.1.