# GPT-Live-1.1 Update Sharpens Multilingual Voice and Search Cues > OpenAI's incremental refresh prioritizes accessibility across languages and interfaces while keeping safety scores steady. *Published 2026-10-16 · By Marcus Vance* GPT-Live-1.1 is OpenAI's incremental successor to the GPT-Live-1 voice model, adding more natural accents in six languages and refining search phrasing while keeping safety metrics stable. OpenAI released GPT-Live-1.1 and the smaller GPT-Live-1.1 Quip on Oct. 15, 2026, positioning the update as an accessibility-first refresh rather than a frontier capability jump. The company said the models are rolling out to Voice in ChatGPT, Codex, and Dots for all users, with GPT-Live-1.1 set as the default voice model for Plus and higher tiers [1][2]. The release follows the GPT-Live-1 model and its mini variant, extending the family with new speech options and search-behavior refinements. The update targets the audible texture of the assistant more than its raw reasoning. New voices are available in French, German, Hindi, Japanese, Korean, and Mexican Spanish, alongside a new English voice named Bubbie [2]. OpenAI said the refresh delivers more natural accents across languages and widens the range of short phrases the model uses during searches, reducing repetitive cues such as 'checking...' [3]. What changed in GPT-Live-1.1? What changed in GPT-Live-1.1? GPT-Live-1.1 is the successor to GPT-Live-1 in OpenAI's live voice-model line, and GPT-Live-1.1 Quip is positioned as a mini variant for lighter workloads [1]. The new release does not introduce a new reasoning architecture or a major benchmark jump; instead, the model focuses on interaction quality in voice channels. In voice-native evaluations using production prompts, GPT-Live-1.1 scored 0.81 on the Sexual category safety metric, compared with 0.77 for GPT-Live-1 [1]. The improvement is small but notable because it came alongside new multilingual voices, suggesting the company tuned safety behavior without sacrificing the broader user experience. | Model | New voice options | Default assignment | Safety score (Sexual) | | --- | --- | --- | --- | | GPT-Live-1.1 | French, German, Hindi, Japanese, Korean, Mexican Spanish, plus Bubbie in English | Plus and higher tiers, including Codex and Dots | 0.81 | | GPT-Live-1.1 Quip | Smaller default model for other users | Non-Plus users | Not disclosed | | GPT-Live-1 | Previous generation voices | Prior default voice model | 0.77 | The comparison below reflects the disclosed specifications in OpenAI's GPT-Live-1.1 system card [1]. Only one safety category metric was published for the voice-native evaluation, so the table does not attempt to rank the models on overall safety. How are the new voices rolling out across ChatGPT, Codex, and Dots? The rollout spans three surfaces: Voice in ChatGPT, Codex, and Dots. OpenAI's announcement on X said the update is rolling out to all users and that API support is coming soon [3]. The system card adds that GPT-Live-1.1 is the default voice model for Plus and higher tier users, including in Codex and Dots [1]. - OpenAI released GPT-Live-1.1 and GPT-Live-1.1 Quip on Oct. 15, 2026. - New voices arrive in Voice in ChatGPT, Codex, and Dots for all users. - GPT-Live-1.1 becomes the default voice model for Plus and higher tiers, including Codex and Dots. - GPT-Live-1.1 Quip serves as the smaller default model for other users. - API support for GPT-Live-1.1 is listed as coming soon. The two-model structure gives OpenAI a way to manage cost and quality across its user base. Plus and higher tiers receive the full GPT-Live-1.1 experience, while the Quip variant keeps the entry-level experience functional without requiring the larger model on every request [1]. What technical details underpin the update? The system card describes GPT-Live-1.1 and GPT-Live-1.1 Quip as the most recent iterations in the GPT-Live model family [1]. The technical changes in this release are concentrated in speech output: voice identity, accent modeling, and the lexical choices the model makes while it works through a search or tool call. OpenAI's release notes describe the update as including more natural accents across languages and new Live voice options for French, German, Hindi, Japanese, Korean, and Mexican Spanish, plus Bubbie in English [2]. The note that the model produces 'less checking...' suggests the speech layer was retrained or reprompted to vary filler phrases rather than default to a single stock utterance [3]. The accent work matters technically because multilingual voice models must map phonemes and prosody from the underlying language into a natural-sounding output. Broadening the voice catalog in six languages at once is a nontrivial data and evaluation effort, even when the underlying language model is unchanged. One distinction in this release is that the speech-layer updates apply to the model's output in both conversational and search contexts. The change from 'checking...' to a wider range of short phrases is designed to make the assistant feel less robotic while it performs a tool call, according to the announcement [3]. Because the same model powers Codex and Dots, the behavior carries across coding and productivity contexts, not just chat. What do the safety benchmarks show? OpenAI's deployment safety documentation includes a limited set of metrics for the voice-native evaluation. The Sexual category score moved from 0.77 on GPT-Live-1 to 0.81 on GPT-Live-1.1, a four-point gain on the same evaluation protocol [1]. The company did not disclose corresponding scores for GPT-Live-1.1 Quip in the cited documentation. The steady-to-improved safety picture matters because GPT-Live-1.1 ships with new voices that expand the model's exposure to non-English speech. OpenAI said the update includes more natural accents across languages [2], which can shift how safety classifiers behave on voice inputs. The benchmark suggests the company preserved, and slightly improved, safety performance while changing the audible surface of the model. For users, the practical effect is that the model sounds different without introducing new safety trade-offs that OpenAI has disclosed. The company's decision to publish even one voice-native safety metric is notable, because voice-specific red-teaming is still less standardized than text-based safety reporting. What role does GPT-Live-1.1 Quip play? GPT-Live-1.1 Quip is the mini version of the model and the smaller default for users outside the Plus and higher tiers [1]. By keeping Quip as the default for other users, OpenAI can deliver the new voice experience to a broad audience while reserving the full model for paying tiers. The Quip naming continues OpenAI's practice of giving lightweight models distinct identities. The system card does not include a separate safety score for Quip, so it is not possible to compare the mini variant with the full model on the disclosed voice-native metric [1]. What do the changes mean for developers and enterprises? For developers, the headline is timing. API support for GPT-Live-1.1 is coming soon, meaning application builders cannot immediately wire the new voices into custom products [3]. When the API lands, teams will likely use the model for voice agents, customer-service bots, and real-time translation flows where filler phrases such as 'checking...' erode perceived quality. Enterprises running ChatGPT Voice in Plus workspaces get the update automatically, since GPT-Live-1.1 becomes the default voice model for higher tiers [1]. That default behavior reduces configuration overhead for teams that rely on the hosted assistant, but it also means organizations may see voice-output style change without an explicit migration step. The multilingual emphasis positions the release as an accessibility play. French, German, Hindi, Japanese, Korean, and Mexican Spanish cover large geographies where English-accented synthetic voices have historically felt less natural. Bubbie, the new English voice, adds a distinct character option for English speakers [2]. How does this rollout compare with OpenAI's release cadence? GPT-Live-1.1 arrives as an incremental version bump rather than a new foundation model. That cadence fits OpenAI's pattern of shipping smaller updates between major model launches, letting the company iterate on behavior without forcing users through a full migration. The simultaneous release of a full model and a mini variant also mirrors the approach OpenAI has used with earlier GPT-Live generations, where a smaller model handles lower-cost use cases [1]. The October 2026 timing puts the update ahead of the year-end voice-assistant buying season, which may explain the focus on multilingual accents and reduced filler phrases. How are experts and users reacting? OpenAI framed the release in plain language. 'New voices. Better accents. And less OpenAI framed the release in plain language. 'New voices. Better accents. And less 'checking...' GPT-Live-1.1 brings these updates to Voice in ChatGPT, Codex, and dots, with new voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese. Rolling out today to all users. Coming soon to the API.' the company said in its announcement on X [3]. New voices. Better accents. And less 'checking...' GPT-Live-1.1 brings these updates to Voice in ChatGPT, Codex, and dots, with new voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese. The announcement treats the release as a user-facing quality improvement rather than a research breakthrough. That framing is consistent with the benchmark data: the model improved on one disclosed safety category and did not introduce a new capability tier. For voice-heavy users, the difference is likely to be felt in long sessions, where repeated filler phrases become noticeable. Some users may notice the change immediately on the first search-enabled voice request, while others will hear the difference only across languages. The 'less checking...' phrasing change, small as it sounds, addresses one of the most common user complaints about AI voice agents: the audible dead air during tool calls. What should users expect on first interaction? On first interaction, a Plus user in any supported language may hear a new voice before they hear a new capability. OpenAI's announcement lists the six language voices and Bubbie as the visible changes, with the search-cue refinement operating underneath [3]. For users who rely on ChatGPT Voice in Codex or Dots, the same model default applies, so the voice change should be consistent across surfaces [1]. What comes next for the GPT-Live family? The near-term roadmap is straightforward: finish the user rollout, ship API access, and let the smaller Quip model find its role as the default for non-Plus users [1]. The split between a full model and a mini variant mirrors OpenAI's broader pattern of pairing frontier models with lighter, cheaper options. Longer term, the release signals that OpenAI views multilingual voice naturalness as a competitive surface. Rival voice assistants have invested heavily in accent fidelity, and GPT-Live-1.1 narrows that gap while keeping the rest of the ChatGPT stack intact. The new language voices give the company a direct answer to the question of whether ChatGPT can sound local, not just accurate. No additional capability benchmarks were disclosed in the rollout materials. The lack of new benchmark numbers is itself informative for an incremental release. OpenAI appears to be measuring success here in interaction polish and safety stability, not in raw model intelligence, which suggests the GPT-Live line will continue to evolve through smaller, more frequent updates rather than monolithic releases. ## Sources 1. [GPT-Live-1.1 is the default voice model for Plus and higher tier users, including for Codex and Dots; GPT-Live-1.1 and GPT-Live-1.1 Quip are the most recent iterations in the GPT-Live model family.](https://deploymentsafety.openai.com/gpt-live-1-1) 2. [New Live voices include options for French, German, Hindi, Japanese, Korean, and Mexican Spanish, plus Bubbie in English, with more natural accents across languages.](https://openai.com/sw-KE/products/release-notes/) 3. [GPT-Live-1.1 brings new voices and less 'checking...' to Voice in ChatGPT, Codex, and dots, with API support coming soon.](https://x.com/OpenAI/status/2110797830361436604) --- Source: https://aiintelreport.com/frontier-models/gpt-live-1-1-multilingual-voice-update Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt