GPT-Live-1.1: 13 new voices, better accents, and the smaller Quip model

What OpenAI announced on 15 October 2026

The announcement went up in OpenAI's community forum on 15 October 2026 under the heading "New voices. Better accents. And less 'checking…'", and the same three claims were repeated on OpenAI's social channels and in a short video. In OpenAI's words, GPT-Live-1.1 "is an update for ChatGPT, Codex, and dots, introducing 13 new voices and considerably better performance in French, German, Hindi, Mexican Spanish, Korean, Japanese, and many other non-English languages".

So there are three changes, and only one of them is about the voice list:

• 13 new voices across seven languages — the full list is in the next section.
• Better accents and non-English performance in the six languages OpenAI names, plus "many other" languages it does not name.
• Less "checking…" — the short phrase Live speaks while it looks something up now varies instead of repeating, which is the change a daily user is most likely to notice in an ordinary session.

OpenAI also says delegation is now more natural and that steerability has improved, and adds that "continuous improvements to voice models and the overall product experience will become the norm going forward". The rollout is to all users on ChatGPT, Codex and dots, with the API following later.

Read the shape of the release, because it sets expectations: there is no benchmark table and no measured number anywhere in it. Compare that with the Grok Voice 3.0 launch two weeks earlier, which arrived with a full third-party benchmark table — different vendors choose different evidence. For GPT-Live-1.1 the claim to hold OpenAI to is the one it made: better accents, more natural delegation, better steerability. None of it is quantified yet.

The 13 new voices, by language

Before this update the Live voice lineup was nine English-sounding standard voices — Arbor, Breeze and Cove among them. This wave is the first set that is language-specific rather than an English voice tinted with a foreign accent, and it lands in seven languages:

OpenAI's own line is that "the voices you see may vary by region and account", so the table above is what the release notes list, not a promise about your account menu. You pick a voice under Settings → Voice → Voice; if the new names are not there yet, the rollout has not reached that account.

One wording correction worth making, because it circulates: this is 13 new voices, not 13 new accents. The accent improvement is a separate, unquantified claim that applies across languages — including languages that receive no new voice at all. Two different things, one of which you can count and one you cannot.

GPT-Live-1.1 or Quip: which model you get

OpenAI shipped two models, not one. GPT-Live-1.1 is the full version; GPT-Live-1.1 Quip is a smaller version of the same model, and the deployment-safety system card lists the pair together as the most recent iterations in the GPT-Live family. Which one answers you depends on your plan, and the plan also sets how many hours of voice you get per rolling 24 hours:

Read this table as second-hand. Those limits come from OpenAI's ChatGPT Voice help page as reported by independent write-ups on 16 October; when this page tried to load the help centre the request was refused (HTTP 403, Cloudflare challenge), so the numbers were not read at first hand. The reporting also says the documents do not state whether this update changed any of the quotas. Check the Voice section of your own account and OpenAI's help page before you plan around a number. For the developer-side rates, use the figures this site keeps current on the GPT-Live API page and the pricing page rather than any plan table copied into a news story.

What is not published is the thing you would most want to know: how much quality Quip gives up. OpenAI describes it as smaller and leaves it there. If you are on Free or Go, that means the model answering you after this rollout is a different, lighter model, and no public number compares the two.

What "better accents" does and does not mean

The accent claim covers French, German, Hindi, Mexican Spanish, Korean and Japanese, plus "many other non-English languages" that OpenAI does not name. That is the whole published scope. There is no per-language score, no accent benchmark, no before-and-after sample set and no list for the unnamed rest — so the claim is real but unfalsifiable from outside, and anyone selling you a language ranking off the back of this update is inventing it.

What the update does not change is how you tell ChatGPT which language you speak. That is still the Settings → Voice → Language control, and recognition quality still depends on the microphone, the room and how much of the sentence is product names and numbers. Better accents on the output side do not upgrade recognition on the input side, and they say nothing about translation between two languages you both speak.

The six named languages were previously recorded on this site as present but unverified with no published roadmap; see the GPT-Live languages page for the setting-level detail and the tier caveats, which this news page does not repeat. That page is about choosing and testing a language pair; this one is about what changed on 15 October.

How to check whether your account has it

The rollout is gradual and OpenAI has not published an end date, so a missing voice is not evidence of a broken account:

• Update the app first on iOS or Android; the new voice names appear in the settings menus, not in the conversation.
• Settings → Voice → Voice — if you can see names such as Bubbie, Sonna or Hanul, the update has reached you. The list depends on your region and account.
• Settings → Voice → Language — confirm the language you speak most, separately from the voice you pick.
• Settings → Voice → mode — Live, Advanced and Standard are different experiences; a changed voice in one mode is not proof that the other modes changed.

Then test the claim instead of trusting it. Read the same 30-second passage in the language you care about with two different voices; repeat it once with background noise; include a name, a date and a number and listen for whether they survive; and ask a question that forces a search so you hear the new filler phrases while it works. That is a fifteen-minute check that tells you more about your own use case than any announcement can.

For developers: the API is still on the earlier Live model

OpenAI's announcement says the update is "coming soon to the API" and gives no date, and the developer documentation still shows the earlier Live model in its examples. Practically: the voices and accents described on this page are a ChatGPT-surface feature today, and a voice agent you build keeps the model, the voices and the behaviour it had before the rollout.

Part of the build-side change did ship with the announcement, though. OpenAI published a guide for native delegation through the Decisions API: instead of Live notifying your application and your application calling Decisions, Live calls Decisions directly, which removes a round trip while the user is still speaking. You declare it when you create the session — a delegation object with the choices you want decided — and read the answer back from the decision event, matching each answer to its question by name. The questions can be changed mid-session, and the delegation type is fixed for the session once set.

What that changes for a voice agent is the timing question, not the voice question: an action can be chosen from a partial utterance rather than after the turn ends. What it does not change is your responsibility for the action itself — your application still performs it, and still has to report back in a way the model can turn into an ordinary sentence. The model, the endpoint and the voice-layer price on this site's API guide remain the ones to build against until OpenAI adds 1.1.

What users reported in the first 48 hours

The announcement thread in OpenAI's community forum filled up with the other half of a voice release: existing voices that do not sound the way they used to. Within a day of the rollout, users wrote that Sol had dropped in pitch and lost warmth, that Spruce no longer matched its old description, and that Arbor sounded different in the paid voice modes; one reported that a saved preference for an American accent still drifted on some turns.

Those are individual reports in a public thread, not OpenAI statements, and nobody on this page listened to any of them. Treat them as a sample of reactions rather than a measurement — the same thread contains users who could hear no difference. It is on this page because it is the practical risk of a voice update: if a workflow, a demo or an accessibility setup depends on one specific voice, re-record it before your audience hears the change for you.

What this page could not verify

• We have not used it. No account was signed in for this page, no audio was compared and no API call was made. Everything above is sourced from published announcements, documentation and reporting.
• The help centre was unreadable to us. OpenAI's release notes and Voice FAQ returned HTTP 403 to this environment, so the model-per-plan mapping and the Voice-hour figures were read through independent write-ups quoting that page, not from the page itself.
• No numbers were published. Accent quality, quality versus Quip, latency and search-cue behaviour all come from OpenAI's own prose with no measurement attached, and no third party has published a re-run.
• "Many other non-English languages" is unnamed. Dutch, Polish, Arabic and everything else are simply not addressed; not being named is not evidence of improvement or of neglect.
• The voice list is the release-notes list, not your selector. OpenAI states availability varies by region and account, and this page cannot enumerate your account.
• A commercial-use pointer we could not confirm. One write-up cites section 8 of OpenAI's service terms as limiting Voice output to non-commercial use and barring redistribution as standalone audio. That would matter to anyone recording a voice for a video or a podcast, but openai.com blocked this page's fetcher, so the clause text was not read here. Check the current terms yourself.

Sources

• OpenAI community forum — GPT-Live-1.1 is rolling out with new voices and better accents: the 15 October announcement, its wording on the 13 voices, the named languages, delegation and steerability, the "coming soon to the API" line, and the user replies quoted under early reports.
• OpenAI deployment safety hub — GPT-Live-1.1 system card: names GPT-Live-1.1 and GPT-Live-1.1 Quip as the current iterations in the family, and documents the intended-use and disallowed-content evaluations for voice.
• OpenAI developer docs — Connect voice to Decisions: native delegation, the delegation session object, question naming, decision events, and the mid-session update rule.
• OpenAI help centre — ChatGPT release notes: the 15 October entry itself (could not be loaded from this environment, HTTP 403; cited as the source the coverage below quotes).
• The Rundown — GPT-Live-1.1: what's new with ChatGPT Voice: the change table, the voice list per language, the note that the release contains no measured speed or accuracy gain, and the service-terms pointer.
• Notebookcheck — GPT-Live-1.1 brings new voices for six languages: the plan-by-plan Voice hours read from the Voice FAQ on 16 October, the previous nine-voice lineup, and the gradual-rollout note.
• DataNorth — OpenAI GPT-Live-1.1: new voices for ChatGPT Voice: the Free/Go-versus-paid model split, the settings paths, the 30-day audio-retention note, and the API-still-on-1 statement.

Related on this site