Skip to content
NewsProduct launch

OpenAI's GPT-Live now reacts before a user stops talking and builds voices from text alone

· by Pondero Newsdesk

The short version

OpenAI shipped two GPT-Live API updates in mid-October 2026: a native Decisions integration that lets an app act mid-sentence, and a prompt-to-voice tool that creates a usable voice from a text description with no recording required.

OpenAI's GPT-Live now reacts before a user stops talking and builds voices from text alone

OpenAI shipped two GPT-Live API updates in mid-October 2026: a native Decisions integration that lets an app act or react while a caller is still mid-sentence, and a prompt-to-voice tool that builds a usable voice from a text description alone, per OpenAI's API changelog.

What

The changelog entry dated October 15 shows GPT-Live gaining native Decisions API integration. Previously, GPT-Live had to notify a developer's own backend mid-conversation and wait for that backend to call the Decisions API itself. Now GPT-Live can call Decisions directly, skipping that round trip, per OpenAI's documentation for the feature. Developers attach triggers to transcript events, such as session.input_transcript.delta while the caller is still speaking or session.output_transcript.delta as GPT-Live itself responds, and each trigger returns a short set of choices, like an expression, an intensity level, or a gesture.

The following day's entry, October 16, added prompt-to-voice: a way to create a custom GPT-Live voice from a written description rather than a recorded sample. A developer sends a text prompt, such as a description of a calm, warm support-line voice, to the /v1/audio/voices endpoint or the voice Playground, and OpenAI generates and saves a reusable voice ID. Creating the voice is free; using it to generate speech inside a GPT-Live session bills at regular session rates, per OpenAI's documentation. Prompt-generated voices work only inside GPT-Live, not the separate Text to Speech, Realtime, or Chat Completions audio APIs.

A round trip disappears, and so does the recording session

Native Decisions integration removes a server hop that previously sat between every delegated decision and GPT-Live's response, which should narrow the gap between a user's utterance and a visible or audible reaction in latency-sensitive builds, such as responsive avatars or in-car assistants. Prompt-to-voice removes a different constraint. Building a distinct voice used to mean either picking from OpenAI's existing presets or supplying a consent recording and a reference sample from a real speaker. A team can now describe a voice in a sentence instead of running a recording session and clearing consent paperwork first. OpenAI's usage policy still requires telling listeners the speech is AI-generated and bars voices modeled on real individuals without consent, so the compliance step moves rather than disappears.

This is a narrower, voice-specific sibling to the Decisions API itself, which OpenAI opened in beta on October 6 as a general-purpose classification endpoint; Pondero's earlier coverage of that launch has the pricing and latency detail for the standalone API.

What to watch next

Watch whether third-party voice-agent builders on GPT-Live start shipping distinct brand voices now that creation no longer needs a recorded sample, and whether OpenAI extends prompt-to-voice to the Text to Speech, Realtime, or Chat Completions audio APIs; OpenAI's documentation lists none of those three as supporting it today.

Sources