Source: Build more natural voice experiences with GPT‑Live‑1 in the API

Voice assistants already sound fairly natural, but many still handle pauses and interruptions poorly. Pause briefly and they may jump in; interrupt them and they may not stop.

GPT-Live-1 is designed to address these problems. It processes incoming and outgoing audio in one model and lets developers delegate deeper reasoning and tool calls to a backend model.

Architecture

A common voice-agent stack has three stages:

Speech-to-text STT → language model LLM → text-to-speech TTS

Every handoff adds latency. When a user pauses, interrupts, or changes direction, the system must also decide whether the turn is over. That is where the conversation often breaks down.

GPT-Live-1 puts the voice layer into one model:

GPT-Live-1: listens + speaks + manages turns and interruptions
                              ↓
             Backend model: deeper reasoning + tool calls + actions

GPT-Live-1 handles listening, speaking, turns, and interruptions. A backend model handles retrieval, reasoning, coding, or actions in business systems. Complex tasks can go to GPT-6 Astra, another text model, or a third-party model.

This setup lets developers control voice latency, backend capability, and cost separately. The voice model does not have to perform every reasoning task itself.

Pauses, interruptions, and silence

GPT-Live-1 processes incoming and outgoing audio together. Users can pause, interrupt, correct themselves, laugh, or give a short acknowledgment without waiting for the agent to finish. The system can also wait through a short pause instead of jumping in or narrating every background step.

A product can define when to reply, when to wait, and when to give only a brief acknowledgment. Language learning, interviews, and medical scheduling all depend on this timing. Natural-sounding audio alone is not enough.

Explicit turn boundaries are still available. GPT-Live-1 supports turn detection, provides ASR transcripts and response text, and supports alphanumeric recognition and keyword biasing.

Developers can use the system prompt to control tone, pace, and conversational style. GPT-Live-1 also adds more voice options across accents, dialects, and languages. Custom voices require contacting OpenAI sales and going through the eligibility and request process.

OpenAI also lists better long-session reliability and telephony support, including reservation, order, and customer-support workflows. These capabilities still need to be tested in real deployment conditions.

Metrics from the release materials

  • According to OpenAI’s evaluations, GPT-Live-1 improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1, mainly in turn-taking latency and interactive behavior.
  • Paired with GPT-6 Astra at medium reasoning effort, it ranks first on Tau3, an end-to-end voice-agent benchmark.
  • In Speak’s early evaluations, interruptions during learners’ thinking pauses fell by almost 80% compared with its previous turn-based systems.
  • A medical voice-product team said that moving from a cascaded architecture reduced its codebase by about 80% and removed roughly 23,000 lines of code.

These numbers come from OpenAI and its customers. I have not independently verified them, so I do not treat them as proven product results.

Cost

GPT-Live-1 is available in the API. The front-end voice layer costs $0.05 per minute, or about $3 per hour of continuous use.

A working product still needs a backend model and an agent harness. The backend model and the number of calls add to the bill. Cost estimates should cover a complete task, not just the per-minute voice price.