OpenAI rolls out GPT-Live-1.1 with new voices and better accents in ChatGPT

What OpenAI shipped on 15 October

OpenAI documented the change at the top of its ChatGPT release notes on 15 October 2026. The note says the company is rolling out GPT-Live-1.1 and GPT-Live-1.1 Quip in ChatGPT Voice, with more natural accents across languages, and that Live now uses a wider range of short phrases to signal when it is searching.

The same day, an OpenAI community post titled GPT-Live-1.1 is rolling out with new voices and better accents repeated the language list and added that the update reaches Voice in ChatGPT, Codex and Dots, is rolling out to all users, and that API access is coming soon. That post is the clearest public statement tying Codex and Dots to the same model family on day one.

OpenAI also published a GPT-Live-1.1 system card dated 15 October 2026 on its Deployment Safety Hub. The card positions GPT-Live-1.1 as the successor to GPT-Live-1 and Quip as the successor to GPT-Live-1 mini, and says GPT-Live-1.1 is the default voice model for Plus and higher tiers, including Codex and Dots, while Quip is the default for everyone else.

What GPT-Live-1.1 and Quip actually are

According to the system card, both models are full-duplex voice systems: they can listen and respond continuously instead of waiting for a clean turn boundary. That is the same design OpenAI introduced with GPT-Live-1. The models decide in the moment whether to speak or keep listening when a user pauses, interrupts or changes pace.

The ChatGPT Voice help article was updated to name GPT-Live-1.1 and Quip as the engines behind the Live option. Live can use web search and memory, show visual results through supported widgets, work with text and images in the same chat, and use plugins and connected apps available to the account. Live still does not support video or screen sharing; those remain on Advanced Voice for eligible subscribers.

Quip is described as a mini version of the model. OpenAI does not publish parameter counts for either checkpoint in the help centre or the system card. The practical distinction for users is which default they get: Plus and above get GPT-Live-1.1; Free and Go get Quip, with plan-specific hour caps.

New voices and where they appear

The release notes list new Live voices for French, German, Hindi, Japanese, Korean and Mexican Spanish, plus Bubbie in English. The Voice help article expands that into named pairs: Aube and Rivage (French), Sonna and Miko (German), Leher and Neer (Hindi), Himari and Rindo (Japanese), Haetsal and Hanul (Korean), Tinta and Arrullo (Mexican Spanish), Viola and Rio (Brazil), and Bubbie (English), alongside the existing English set such as Arbor, Breeze, Cove, Ember, Juniper, Maple, Sol, Spruce and Vale.

OpenAI is explicit that the voices you see may vary by region and account. Changing the selected voice during a Live conversation starts a new voice call in the same chat. Preset ChatGPT personalities still do not apply to Live; users can still ask for a different tone or pace inside a single call.

The community announcement frames the update as fewer generic 'checking…' moments as well as better accents. That matches the release-notes claim that Live uses a wider range of short phrases while searching. Neither document publishes a measured reduction in filler phrases; both treat it as a product behaviour change.

How Live fits ChatGPT Voice options and plan limits

ChatGPT Voice still offers three modes under Settings → Voice: Live, Advanced and Standard. The help article says Live is the full-duplex GPT-Live experience; Advanced keeps video and screen sharing for eligible mobile subscribers; Standard remains turn-by-turn speech that transcribes before answering. For broader ChatGPT model context, see our earlier note on GPT-6 and Intelligent UI in ChatGPT.

Voice in Chat usage is measured over a rolling 24-hour window. The help article's table, updated for GPT-Live-1.1, lists Free as limited Quip access; Go as three hours of Quip; Plus as three hours of GPT-Live-1.1; Pro at $100 as fifteen hours; Pro at $200 as unlimited subject to abuse guardrails; Business Standard as three hours plus 1.25 credits per extra minute; Business Premium as fifteen hours plus the same overage; K-12 Teachers as three hours; credit-based Enterprise, Edu and Clinicians as 1.25 credits per minute; usage-based Enterprise as $0.05 per minute; and legacy Enterprise and Edu as three hours.

Voice in Work follows the same Voice limits and also counts toward normal Work usage. Desktop Voice has separate pricing and limits. Workspace settings still gate Business, Enterprise, Edu and Healthcare access.

What the system card says about safety

The system card says OpenAI built safeguards for GPT-Live-1.1 largely the same way as for GPT-Live-1: train the models to respond safely, measure voice-specific evaluations, reuse system-level safety integrations, and keep the same monitoring and enforcement stack.

On voice-native evaluations with production prompts, OpenAI reports the new models at parity with their predecessors and says there were no statistically significant regressions. On synthetic prompts built to hit hard safety edges, the card says performance is close to parity, but GPT-Live-1.1 shows statistically significant regressions versus GPT-Live-1 on sexual, mental-health, emotional-reliance and gore categories. OpenAI says more research is needed and that it plans to improve those categories in later releases. Those synthetic scores also omit system-level tools such as Trusted Contact, crisis helplines and parental controls.

Alignment tests converted to audio show mixed small shifts: slight regressions on respecting environmental warnings, a slight improvement for GPT-Live-1.1 and a slight regression for Quip on blocker deception, and the opposite pattern on a broken-search-tool disclosure test. The card argues that Live models inherit alignment behaviour from their underlying thinking models and that neither Live checkpoint, without delegation, plausibly reaches High under OpenAI's Preparedness Framework for bio, cyber or self-improvement.

Codex, Dots and the API timeline

The system card's defaulting rule matters for product surfaces beyond consumer Chat: Plus and higher tiers get GPT-Live-1.1 for Codex and Dots as well as ChatGPT Voice. The community announcement is the source that says the rollout includes those surfaces on day one and that the API is still forthcoming.

OpenAI has not published API model IDs, pricing or a date for Live 1.1 in the developer docs as of this write-up. Teams that already use Realtime or speech APIs should treat the community 'coming soon' line as a product promise, not a shipping schedule.

For readers comparing speech stacks, our coverage of xAI's Grok Voice 3.0 API and Fish Audio Drama 3 shows how differently vendors package speech-to-speech: some lead with API pricing and benchmarks, while OpenAI's Live 1.1 day-one story is still centred on ChatGPT surfaces and a safety card.

How this compares with other speech models

GPT-Live-1.1 is not a greenfield launch. It is an incremental model family update inside an already shipping Live experience. That makes the interesting deltas accents, voice inventory, search-status phrasing and the published safety deltas, not a claim that OpenAI invented full-duplex voice this week.

OpenAI still separates Live from Advanced. If you need the camera or screen-share path on mobile, you stay on Advanced. If you only need spoken back-and-forth with search, memory, widgets and plugins, Live is the path the company is actively updating.

The company also keeps Standard as a turn-by-turn option that deletes audio after transcription unless the user opts into sharing clips for training. Live and Advanced retain audio and video clips with the chat transcript for 30 days under the help article's retention rules.

Limits and open questions

OpenAI has not published independent accent-quality scores, a full voice catalogue by country, or API availability details for GPT-Live-1.1. Regional voice lists can differ from the help article's table, and the community post's 'all users' line sits next to plan and workspace gates that still apply.

The system card's synthetic-prompt regressions are OpenAI's own measurements. They are not prevalence-weighted production rates, and the company says system-level interventions are outside that table. Until API docs land, developers should not assume Live 1.1 pricing or latency from ChatGPT plan limits.