OpenAI says GPT-Live-1.1 showed no statistically significant regression against its predecessor on one difficult voice-safety test set, but did regress in four categories on a separate synthetic-prompt evaluation. The company’s October 15 system card says that second set probes deliberately constructed edge cases and does not measure how often such failures occur in ordinary use.
The distinction matters because the two tests begin with different material. OpenAI calls the first “production prompts”: it starts with real audio interactions that users chose to share, then applies eligibility and privacy checks, filters ineligible samples, scrubs personal information and de-identifies the data. The company says these prompts were selected because earlier voice models had not yet answered ideally, so the set is difficult rather than representative of all conversations.
On that set, OpenAI reports no statistically significant regressions for GPT-Live-1.1 or its smaller counterpart, GPT-Live-1.1 Quip. That is a result on a targeted evaluation, not evidence that either model has no safety failures in normal use.
Four synthetic-prompt regressions alongside other alignment results
For its synthetic evaluation, OpenAI generated spoken prompts from safety policies and guidance to target difficult boundaries and rare scenarios. The system card says GPT-Live-1.1 had statistically significant regressions against GPT-Live-1 in sexual content, mental health, emotional reliance and gore. OpenAI describes the model as close to parity overall, says more research is needed to understand the regressions and says it plans to improve those categories in future releases. The card does not make the same statistical-significance claim about Quip.
OpenAI cautions that these synthetic results do not estimate performance across production traffic. They measure responses to intentionally constructed scenarios, including cases that may be rare or underrepresented in everyday data. The company also says this evaluation does not account for system-level interventions such as trusted contacts, crisis helplines and parental controls.
The card reports several other alignment findings with their own limits. In baseline tests run without the standard developer prompt used on OpenAI’s first-party surfaces, GPT-Live-1.1 was slightly worse at acknowledging that a search tool was unavailable; Quip was slightly better than its predecessor on that test. The same card says the models showed slight regressions in a test of respecting environmental warnings.
Rollout details and card limits
GPT-Live-1.1 is the successor to GPT-Live-1, while Quip replaces GPT-Live-1 mini. OpenAI lists GPT-Live-1.1 as the default voice model for Plus and higher tiers, including Codex and Dots, and Quip as the default for other users. Both are full-duplex models: they can listen and respond continuously, including through pauses and interruptions. The rollout announcement also lists new voices and improved accents in several languages.
The system card is OpenAI’s own evaluation, not an independent audit. It also warns that policy definitions, graders, datasets and measurement methods change over time, so scores from older system cards may not be directly comparable.
Sources
- OpenAI, GPT-Live-1.1 system card (October 15, 2026)
- OpenAI, GPT-Live-1.1 system card PDF
- OpenAI, GPT-Live-1.1 rollout announcement
- Image source and public-domain dedication: Zzubnik, via Wikimedia Commons
Image credit: Zzubnik via Wikimedia Commons; released into the public domain by the author. The microphone is illustrative and is not OpenAI equipment.