GPT-Live-1.1 brings new ChatGPT voices and accents; safety card shows mixed results

The voice update

OpenAI's 15 October 2026 release notes say GPT-Live-1.1 and its smaller companion GPT-Live-1.1 Quip are rolling out in ChatGPT Voice. The change adds voice options for French, German, Hindi, Japanese, Korean and Mexican Spanish, plus an English voice called Bubbie. OpenAI says Live now uses a wider range of short phrases while searching. Voice availability can differ by account and region.

The system card identifies 1.1 as the successor to GPT-Live-1 and Quip as the successor to GPT-Live-1 mini. These are full-duplex models: they can listen and speak as a conversation unfolds rather than requiring rigid alternating turns. That supports pauses, interruptions and more natural timing; it does not itself establish lower latency or higher accuracy without a benchmark.

Which model and voice will a user see?

The model names describe the engine used for the Live experience, while voice names describe how speech sounds. Selecting a voice in Settings does not mean a user is choosing a different reasoning model. OpenAI says 1.1 is the default for Plus and higher tier users, including Codex and Dots, while Quip serves other users. The release notes also warn that voice choices vary by account and region.

The new accent and voice choices address a real usability issue: a multilingual assistant is easier to understand when its pronunciation and timing suit the language being spoken. The release notes do not give a numerical accent-quality measure or a per-language error-rate table, so 'better accents' remains OpenAI's product assessment rather than an independently quantified result.

The model and interface distinctions stated by OpenAI on 15 October 2026.

Item	What OpenAI says	What it does not mean
GPT-Live-1.1	Successor to GPT-Live-1; default for Plus and higher, including Codex	Every account has the same voices or limits
GPT-Live-1.1 Quip	Smaller successor to GPT-Live-1 mini; default for other users	Identical capability to the larger model
New voices and accents	Options for French, German, Hindi, Japanese, Korean, Mexican Spanish and English Bubbie	A published numerical benchmark for each language
Full-duplex conversation	Can listen and respond continuously, accounting for pauses and interruptions	A guarantee that every interruption is understood correctly

The system card's safety results need two readings

OpenAI compared each 1.1 model with its predecessor on two voice-native safety sets. One uses production-derived prompts from users who chose to share audio, processed through eligibility, privacy and de-identification steps. OpenAI reports parity with no statistically significant regressions on this set. The other uses synthetic audio prompts designed to target difficult policy boundaries. On that set, OpenAI reports statistically significant regressions for GPT-Live-1.1 in sexual content, mental health, emotional reliance and gore categories compared with GPT-Live-1. The card says more research is needed and that it plans improvement.

Both sets were selected to be difficult and are not prevalence weighted. They do not tell us the frequency of unsafe responses among all real conversations. The synthetic results also exclude some system-level interventions, including crisis resources, Trusted Contact and parental controls. That context neither erases the regressions nor turns them into a real-world incident rate.

Same launch, different evaluation designs and conclusions; do not combine the rows into a single safety score.

Evaluation	Prompt origin	OpenAI's reported finding	Important limit
Production-derived voice-native	Shared real audio, processed for privacy and eligibility	1.1 models at parity with predecessors; no significant regressions	Challenging subset, not live-traffic prevalence
Synthetic voice-native	Generated audio aimed at policy edge cases	1.1 near parity overall, with significant regressions in four named categories	Does not include all product-level interventions
Broken-search tool check	Task that normally needs unavailable search	1.1 slightly worse and Quip slightly better than their predecessors	A narrow honesty test, not general voice quality

What this means in use

Users can check Settings → Voice for the new voice choices available to their account. Multilingual users should test names, regional terms, interruptions and short back-and-forth questions in their own language rather than assume that a general accent claim predicts their experience. Developers should not treat this ChatGPT rollout as confirmation of immediate API availability; the release notes here speak to ChatGPT Voice.

For readers assessing model quality, the clearest technical advance described publicly is the updated voice experience and language options. The system card supplies safety comparisons, with the caveats above. OpenAI does not publish a comprehensive independent latency, speech-recognition or multilingual accent benchmark in the two primary sources cited here.

Sources and provenance

The rollout, new voices and selection route come from OpenAI's 15 October release notes. Model tiers, full-duplex operation and evaluation findings come from the GPT-Live-1.1 system card. This article interprets OpenAI's disclosed tests; it has not independently benchmarked the models. Read 16 October 2026.