What GPT-Live-1.1 Changes in ChatGPT Voic

A voice assistant can give the right answer and still be frustrating to talk to. You pause to think, it starts answering, and you have to interrupt to finish your point. OpenAI announced GPT-Live-1.1 on October 15, 2026, adding voice options and what it describes as more natural accents. ChatGPT also varies the brief spoken messages it gives during a search. ChatGPT release notes

For people who use Voice to think through a problem or direct a task, pauses and corrections are part of the conversation. The useful test is whether the assistant follows those changes and carries them through to the work it does.

New voices, familiar full-duplex conversation

GPT-Live-1.1 succeeds GPT-Live-1. Its smaller sibling, GPT-Live-1.1 Quip, replaces GPT-Live-1 mini. Both retain full-duplex operation: they can listen while speaking and respond to changes in conversational pace. That behavior predates this update.

The October release adds voice options for French, German, Hindi, Japanese, Korean, and Mexican Spanish, alongside Bubbie in English. OpenAI says accents sound more natural across languages. The voices available to you may vary by account and region.

For language practice, the new voices give users more choice over what they hear throughout a session. Search updates serve a different purpose: they tell you why a response is taking longer, without making you look at the screen. You can hear that a lookup is underway and continue the conversation when the result arrives.

The voice and the task model have different jobs

OpenAI's developer documentation describes GPT-Live as a conversational layer that can delegate work to a backend model or agent. The backend looks up information and uses tools while Live handles the spoken exchange. Developers choose the backend independently and can connect either a supported Responses model or their own service. Getting started with GPT-Live

Consider a hypothetical order-status assistant. A customer asks where a parcel is, then corrects the order number while the lookup is still running. The assistant needs to register the correction, and the application needs to stop or disregard the old lookup before giving a result. If a later request would change the delivery address, the application must also check permissions and obtain any required confirmation.

A useful product test should catch both kinds of failure: an assistant that talks over the customer and one that sounds helpful but checks the wrong order.

In the desktop app, Voice can use an existing Codex task's conversation and selected model, letting you ask about progress or give follow-up instructions. Availability depends on rollout and workspace settings. For a coding task, that means evaluating both the conversation and the code produced by the selected model. ChatGPT desktop Voice

Access and cost depend on where you use Voice

As of October 16, 2026, OpenAI's Voice guide lists the following consumer plan limits for Live in Chat, measured over a rolling 24-hour period. ChatGPT Voice usage limits

Desktop Voice is billed differently. OpenAI lists $0.05 per minute against the existing Codex usage budget, with the task model charged separately at its token rates. The desktop Voice rate is 1.25 credits per minute for Business, Edu, and Enterprise workspaces on credit-based billing, and for Plus and Pro users spending additional credits. Desktop Voice pricing

For developers, API availability is a separate question. As of October 16, OpenAI's API model page lists GPT-Live 1 under the identifier gpt-live-1, at $0.05 per minute, billed per second. Backend model and tool usage is charged separately. That page does not establish a GPT-Live-1.1 API identifier or rate. GPT-Live 1 API model

When estimating costs, count the time spent talking as well as the work done in the background. A long discussion followed by several searches can cost more than a short instruction. The per-minute voice rate covers only part of that bill.

What the safety results establish

OpenAI reports that the new models performed on par with their predecessors in safety evaluations based on selected real user prompts, with no statistically significant regressions. On synthetic audio prompts targeting difficult cases, however, GPT-Live-1.1 showed statistically significant regressions relative to GPT-Live-1 in the sexual content, mental health, emotional reliance, and gore categories. The synthetic tests exclude some system-level protections, and neither set of results measures how often failures occur in everyday use. Published safety findings

These are OpenAI's own evaluations of safety behavior under specific conditions. They do not establish a general improvement in accuracy or usefulness. For anyone assessing a voice product, it is worth testing how the assistant responds when it lacks information or cannot finish a task. A confident delivery can otherwise make a weak answer easy to overlook.

Try a conversation that needs a correction

To get started, select Live under Settings → Voice and choose an available voice. In ChatGPT, Live supports text and images in the same chat when available to your account. Video and screen sharing require Advanced on eligible iOS or Android accounts.

Start in a quiet setting, since OpenAI notes that background noise and overlapping speech can still cause interruptions. A useful first trial is to explain a small planning problem, pause, and then correct one important constraint while the assistant responds. For example, describe a presentation for ten colleagues, then change the audience to first-time customers. Listen for whether the assistant picks up the correction, then check whether the outline actually suits the new audience. Repeat the exercise in the language you expect to use regularly.

GPT-Live-1.1 is intended to make conversations with ChatGPT feel more natural. Try it on a task you know well, and judge the result by whether you can explain what you need, correct a detail, and finish the task without repeatedly starting over.