OpenAI GPT‑Live‑1 API release

TL;DR

OpenAI announced GPT‑Live‑1 in the API, a full‑duplex voice model that can listen and speak at the same time, dramatically improving interruption handling, latency, and conversational naturalness for voice‑enabled applications.

What GPT‑Live‑1 Is

GPT‑Live‑1 is a single model that performs automatic speech recognition (ASR), reasoning, and text‑to‑speech (TTS) in one unified architecture. It was first demonstrated in ChatGPT and is now available to developers via the OpenAI API.

Key capabilities include:

  • Simultaneous listening and speaking – the model processes inbound audio while generating outbound speech, eliminating the need for separate STT‑LLM‑TTS pipelines.
  • Smooth interruption handling – it reasons over incoming and outgoing audio together, reducing interruptions by up to 80% compared to turn‑based systems (as measured by Speak’s early evaluations).
  • Delegation to backend models – deeper reasoning and tool calls can be off‑loaded to a text model such as GPT‑6 Astra or third‑party models.
  • Customizable tone, pace, and style – developers control the agent’s voice characteristics through the system prompt.
  • Robust background‑noise management – the model distinguishes silence and ambient sounds without breaking the conversational flow.
  • Long‑session reliability – it retains context over extended interactions, improving conversational quality.
  • Telephony support – enables full‑duplex agents for phone‑based use cases like reservations and customer support.

Architectural Simplification and Latency Reduction

Traditional voice agents stitch together separate speech‑to‑text, reasoning, and text‑to‑speech components. Each handoff adds latency and introduces points of failure, especially when users interrupt or pause.

GPT‑Live‑1 collapses this stack into a single model, handling:

  1. ASR transcription – providing real‑time text of the user’s speech.
  2. Turn detection – recognizing when a user finishes speaking, even in a non‑turn‑based setting.
  3. Response generation – producing speech while still listening for interruptions.
  4. Tool delegation – forwarding complex queries to a backend model (e.g., Luna for high‑volume tasks, Astra for nuanced reasoning).

By keeping the conversation alive while background reasoning occurs, developers achieve lower end‑to‑end latency and a more natural rhythm.

Measured Full‑Duplex Advantages

OpenAI’s internal evaluations show substantial gains for GPT‑Live‑1:

  • Full Duplex Bench performance improves by 30 percentage points over GPT‑Realtime‑2.1, with notable reductions in turn‑taking latency.
  • When paired with GPT‑6 Astra (medium reasoning effort), GPT‑Live‑1 ranks #1 on the Tau3 metric, which measures frontier voice‑agent intelligence on end‑to‑end tasks.
  • Across airline, retail, telecom, and banking spoken‑customer‑service benchmarks, the model achieves higher Pass@1 scores for task success, interruption handling, and tool‑use accuracy.

These results demonstrate that a unified full‑duplex architecture can outperform the best existing modular pipelines on both speed and task success.

New Voice Options and Language Coverage

GPT‑Live‑1 expands beyond the limited set of real‑time voices previously offered. Developers can now choose from a broader palette of accents, dialects, and languages, starting with an Australian English‑influenced voice. OpenAI plans to continue adding voices and language support in the coming months.

Pricing and Availability

  • Front‑end voice layer: $0.05 per minute.
  • The model is available in the API today; developers can pair it with any backend model or tool harness that fits their cost and performance requirements.
  • Custom voice access and enterprise‑grade deployments are available through sales contact or via OpenAI Presence, which leverages GPT‑Live‑1 for trusted real‑time voice interactions in enterprise settings.

How Developers Can Use GPT‑Live‑1

  1. Start a session – invoke the API, speak naturally, and test interruption handling in noisy environments (e.g., coffee shops, city streets).
  2. Shape the agent – set system prompts to adjust tone, pace, and conversational style.
  3. Delegate reasoning – configure the backend model (e.g., GPT‑6 Astra) for complex queries or tool calls.
  4. Integrate with existing workflows – use the provided ASR transcripts and response text for logging, analytics, or downstream processing.

Customer Feedback Snapshot

Early adopters report that GPT‑Live‑1’s interruption handling allows learners more thinking time before a tutor responds, cutting interruptions by almost 80 % compared to prior turn‑based systems.

Takeaway

GPT‑Live‑1’s release marks a shift from fragmented voice pipelines to a unified, full‑duplex model that delivers lower latency, better interruption handling, and richer customization, opening new possibilities for natural voice‑first applications across consumer and enterprise domains.

Sources