GPT-Live-1.1 explained: new voices and more natural accents in ChatGPT Voice

The answer in brief

GPT-Live-1.1 is OpenAI’s updated voice model for ChatGPT, released on 15 October 2026. It changes how ChatGPT Voice sounds and keeps everything else: the full-duplex design that listens and speaks at the same time, the features and the reasoning model that writes the answers.

• Where it runs: the Live option in ChatGPT Voice on the web, iOS and Android, plus voice in Codex and in dots, OpenAI’s always-on agents.
• What’s new: 13 voices, including native voices for French, German, Hindi, Japanese, Korean and Mexican Spanish; what OpenAI calls “more natural accents across languages”; and a wider range of the short phrases it says while it searches.
• Who gets it: every plan. Free and Go use the smaller GPT-Live-1.1 Quip; Plus includes 3 hours of GPT-Live-1.1 in any 24 hours, and Pro at $200 a month has no fixed limit (US list prices, checked 16 October 2026; UK subscribers pay a localised price in pounds).
• Developers: GPT-Live-1 stays the API model at $0.05 a minute. OpenAI says 1.1 is coming soon to the API and has given no date.
• Evidence: OpenAI has published no benchmarks for 1.1, independent leaderboards still list only GPT-Live-1, and AIWIZ has not tested 1.1 yet. Until someone measures them, the accent gains are OpenAI’s claim.
• Verdict: if you speak one of those six languages, switch to one of its new native voices and judge for yourself. English speakers will notice little beyond Bubbie, the one new English voice.

What is GPT-Live-1.1?

GPT-Live-1.1 (sometimes written GPT Live 1.1) is the second release of GPT-Live, the voice model family OpenAI introduced on 8 July 2026 to replace Advanced Voice Mode as ChatGPT’s default. These are the key facts, checked against OpenAI’s release notes and help centre on 16 October 2026.

What’s new in GPT-Live-1.1 compared with GPT-Live-1?

GPT-Live-1.1 changes how ChatGPT sounds and leaves what it can do unchanged. OpenAI’s release notes and announcement on X list three changes: new voices, more natural accents and more varied phrases while searching.

More natural accents and native voices

Accent was GPT-Live-1’s weakest point outside English. At launch, OpenAI said that in certain languages the model may have a non-native accent or gaps in fluency. When OpenAI demonstrated live translation into Hindi in July, TechCrunch heard a heavy American accent and Hindi that sounded unnatural.

OpenAI says version 1.1 brings “more natural accents across languages”, and it adds native voices for French, German, Hindi, Japanese, Korean and Mexican Spanish, the languages its announcement on X singles out. It has published no measurements to support the claim.

Less “checking…”

While a backend model searches or reasons, Live fills the wait with short spoken cues. GPT-Live-1 tended to repeat the same few in long conversations. GPT-Live-1.1 draws on a wider range. As Xataka points out, searches are no faster; they sound less repetitive.

The announcement on OpenAI’s developer forum, posted by one of its community moderators, adds that delegation “is now more natural and steerability has been improved”. OpenAI’s release notes do not repeat or explain that claim.

What is GPT-Live-1.1 Quip?

GPT-Live-1.1 Quip is the smaller version of GPT-Live-1.1 and replaces GPT-Live-1 mini. ChatGPT uses it on the Free and Go plans. OpenAI has not explained the name or published how Quip differs from the full model in quality or speed.

What are the new ChatGPT voices?

GPT-Live-1.1 adds 13 voices: two native voices for each of six languages and one new English voice, Bubbie. The ChatGPT Voice help article names them as follows.

The original nine voices remain: Arbor, Breeze, Cove, Ember, Juniper, Maple, Sol, Spruce and Vale. The help article also lists two Brazilian voices, Viola and Rio, which the 1.1 announcement did not mention. That makes 24 voices in the current list, though OpenAI warns that the voices you see may vary by region and account.

How do I change ChatGPT’s voice?

1. Open Settings > Voice in the ChatGPT app or on chatgpt.com.
2. Select Voice and choose from the list.
3. To improve recognition, select Language and pick the language you speak most often.

Changing voice during a conversation starts a new voice call in the same chat. You can also ask ChatGPT mid-conversation to switch language, slow down or change its tone.

How does GPT-Live work?

GPT-Live listens and speaks at the same time, and it hands hard questions to a separate reasoning model. OpenAI described both design choices in its July launch post; version 1.1 keeps them.

What does full-duplex mean?

Full-duplex means the model processes your speech continuously while it talks. OpenAI says GPT-Live makes interaction decisions many times per second: whether to speak, keep listening, pause, interrupt or call a tool. That lets it say “mhmm” while you talk, wait while you think and translate as you speak.

Earlier versions of ChatGPT Voice worked differently. The original chained speech-to-text, a language model and text-to-speech. Advanced Voice Mode used one model but still took turns, treating silence as the sign you had finished, so a pause or a passing lorry could make it cut in.

Which model does ChatGPT Voice use to answer questions?

GPT-Live runs the conversation, and a frontier text model does the heavy thinking. When a question needs web search, deeper reasoning or agentic work, Live delegates it and keeps talking until the answer comes back.

At launch the backend was GPT-5.5. Since 9 September 2026, ChatGPT Voice can use GPT-5.6 or GPT-6 Astra, and you choose the model and reasoning effort with the same controls as text chat. Because the two layers are separate, OpenAI can upgrade one without retraining the other.

The “checking…” fix sits in the gap between them. GPT-Live-1.1 changes what the voice says while it waits; how long the wait lasts still depends on the backend model you pick.

ChatGPT Voice Live vs Advanced vs Standard: what’s the difference?

Live is the GPT-Live-1.1 experience and the default for natural conversation. Advanced is the previous real-time voice, and you still need it for video and screen sharing. Standard transcribes your speech first and answers turn by turn. The comparison below comes from the ChatGPT Voice help article and OpenAI’s launch post.

To switch, open Settings > Voice and select Live, Advanced or Standard. Which options you see depends on your plan, workspace settings, region, app version and any parental controls.

Is GPT-Live-1.1 free? Usage limits by plan

Yes, with limits. Free users get limited access to GPT-Live-1.1 Quip, and every paid plan includes a daily allowance measured over a rolling 24 hours. The ChatGPT Voice help article sets these limits for Voice in Chat (checked 16 October 2026).

The dollar amounts in the table are OpenAI’s US list prices. Subscribers in the UK pay in pounds, at a local price shown at checkout, under OpenAI’s multi-currency billing rules; API credits are priced in dollars only.

What happens when you hit the limit?

ChatGPT tells you when you reach a limit, and the voice conversation can end. You can carry on in text or wait for the rolling window to free up time. Since 9 September 2026, Plus and Pro no longer drop to a smaller voice model at the limit.

Voice used inside ChatGPT Work follows the same limits and also counts towards your Work usage. Voice in the desktop app has separate pricing.

Is GPT-Live-1.1 available in the API?

Not yet. OpenAI says GPT-Live-1.1 is coming soon to the API and has given no date. Developers can use GPT-Live-1, which became available in the API on 10 September 2026.

GPT-Live-1 costs $0.05 a minute, billed per second, for the front-end voice layer (US list price, checked 16 October 2026). You pair it with a backend model of your choice for reasoning and tool calls: OpenAI suggests Luna for high-volume tasks such as order updates, GPT-6 Astra for complex issues, or a third-party model. It returns ASR transcripts and response text, supports keyword biasing and turn detection, and works over the phone. The API has its own 12 voices, including Quartz, Meridian and Bossa; custom voices require contacting OpenAI’s sales team.

How does GPT-Live compare with GPT-Realtime?

OpenAI says GPT-Live-1 beats GPT-Realtime-2.1, its earlier real-time voice model, by 30 percentage points on Full Duplex Bench, with large gains in turn-taking speed. Paired with GPT-6 Astra at medium reasoning effort, it ranks first on Tau3, which OpenAI describes as a measure of voice-agent intelligence on end-to-end tasks, according to OpenAI’s API announcement. That task score depends heavily on the reasoning model behind the voice layer, so it says more about the pairing than about how natural the voice sounds.

Has OpenAI published benchmarks for GPT-Live-1.1?

No. The release notes, forum post and announcement on X contained no figures when we checked on 15 October 2026, and OpenAI has not said how Quip differs from the full model, as AlphaSignal noted. Until independent tests appear, the accent gains are OpenAI’s claim.

The published evidence covers GPT-Live-1. In OpenAI’s human evaluations of matched 5 to 10 minute conversations, testers strongly preferred it over Advanced Voice Mode. Speak, a language-learning app, found it cut interruptions during learners’ thinking pauses by almost 80% compared with its previous turn-based system.

Has GPT-Live been tested independently?

Only the first version has. No independent test covers GPT-Live-1.1 yet, so the evidence available is for GPT-Live-1 in the API. On the Artificial Analysis Speech to Speech leaderboard (checked 16 October 2026), GPT-Live-1 paired with GPT-6 Astra at medium effort scores 83.2. That puts it second, behind SpaceXSI’s Grok Voice 3.0 on 86.9 and just ahead of Google’s Gemini 3.8 Live Extended Thinking on 82.6.

The scores pull in different directions. GPT-Live-1 completes the most simulated customer-service tasks, 74.5% on τ-Voice, though Artificial Analysis bases that figure on a single trial where most models get three. Grok Voice 3.0 starts speaking sooner, 0.81 seconds to first audio against 1.34 for GPT-Live-1.

These are developer set-ups, tested through each company’s API. ChatGPT Voice always runs on OpenAI’s own model, so for ChatGPT users the choices that matter are the voice and the backend model, and a leaderboard place says little about how either sounds in your language. For the developer view, our Gemini 3.8 Live guide compares the three models in detail.

Can you use GPT-Live-1.1 in Codex, dots and the UK?

Yes in Codex and in ChatGPT across the UK; dots are more restricted. OpenAI says the 1.1 rollout reaches all users, though the voices on offer may vary by region and account.

Codex

Voice has worked in Codex in the ChatGPT desktop app on macOS and Windows since 23 July 2026, according to the release notes. You can start tasks, check progress and coordinate several agents by talking. GPT-Live-1.1 now powers that voice.

Dots

Dots are OpenAI’s always-on agents, launched on 29 September 2026 and powered by GPT-6 Astra. Each has its own cloud computer, and you can talk to yours on a voice call. At launch they were limited to eligible Pro and Business Premium users aged 18 and over, and Pro access excluded the UK, the European Economic Area and Switzerland. Our OpenAI DevDay 2026 recap covers dots in detail.

The UK

ChatGPT Voice with Live launched globally on web, iOS and Android, so UK users on any plan can use GPT-Live-1.1 within their plan’s limits. The exception is dots on a personal Pro plan, which were not available in the UK at launch.

What can’t GPT-Live-1.1 do?

Version 1.1 leaves GPT-Live’s known gaps in place. The ChatGPT Voice help article lists them:

• No video or screen sharing in Live. Switch to Advanced on iOS or Android for those.
• Built for one-to-one conversation; it may reply when people in the room talk to each other.
• Transcripts are not verbatim and can differ from what was said, especially when speech overlaps.
• ChatGPT’s preset personalities do not apply to Live, and there is no precise playback-speed control.
• Plugin actions that need approval must be tapped on screen; spoken approval is not supported.
• Live cannot fetch files from your ChatGPT Library, though you can attach files manually.
• You can hold one voice conversation at a time.

Intelligent UI, the charts, forms and small tools that GPT-6 can build into a text answer, is not available in Voice either, according to OpenAI’s Intelligent UI help article. Our guide to GPT-6 and Intelligent UI explains what you are missing.

Why does ChatGPT Voice keep interrupting me?

Live can still cut in during long pauses, in noisy places or when it hears another speaker. OpenAI’s suggestions:

1. Use headphones or move somewhere quieter.
2. On an iPhone, open Control Centre during the call, select Mic Mode and turn on Voice Isolation.
3. At the start, tell it to wait, for example: “Wait until I ask you to respond.”

Long silences or background speech can still trigger a reply even after that instruction.

Does ChatGPT record your voice conversations?

Yes. Audio clips from Live and Advanced conversations are stored with the chat transcript and kept for 30 days, according to the ChatGPT Voice help article. Deleting the chat deletes its clips within 30 days, unless OpenAI must keep them for security, safety or legal reasons.

OpenAI does not train on your audio unless you opt in. In personal accounts, that means turning on Improve the model for everyone and then Include your audio recordings under Settings > Data Controls. With the first setting on, OpenAI may still use transcripts for training, depending on your plan and settings.

On safety, Live speaks only in predefined voices, with safeguards against imitating a real person. Since 31 July 2026, supported audio generated with GPT-Live in ChatGPT Voice and the API has carried a SynthID watermark, which OpenAI’s public verification tool can detect.

When did GPT-Live come out? Release history

GPT-Live launched in ChatGPT on 8 July 2026, and GPT-Live-1.1 followed on 15 October 2026. Dates below come from the ChatGPT release notes and OpenAI’s announcements.

Is GPT-Live-1.1 worth trying?

Yes, if you speak French, German, Hindi, Japanese, Korean or Mexican Spanish. Accent was GPT-Live-1’s most obvious flaw in those languages, and 1.1 gives each of them two native voices. Pick one and listen for yourself.

For English speakers the change is smaller: Bubbie, and a voice that repeats itself less while it searches. OpenAI does not claim smarter answers; those still depend on the backend model you choose.

Developers should keep building on GPT-Live-1 and retest when 1.1 reaches the API. With no published benchmarks, run your own checks: long pauses, mixed-language speech, spelled-out codes and a few minutes of background noise.

Expect more releases like this. OpenAI shipped 1.1 just over three months after GPT-Live-1, with Voice updates arriving every few weeks in between, so further point releases look likely. If you are choosing voice tools for your team, talk to AIWIZ.