GPT‑Live‑1 can listen and speak at once, handle interruptions and delegate complex reasoning or actions to a separate backend model.
OpenAI launched GPT‑Live‑1 in its API on September 10, offering developers a voice model that can receive and produce speech simultaneously. The company positions it for conversational applications and business workflows such as telephone reservations and customer support.
The defining feature is full-duplex operation: GPT‑Live‑1 can continue listening while it speaks, which OpenAI says enables smoother handling of interruptions. A single model processes incoming and outgoing audio instead of relying on a sequential chain of speech recognition, reasoning and speech synthesis components.
Developers can use prompts to control speaking tone and pace. OpenAI also reports better handling of silence and background noise, improved reliability in longer sessions, alphanumeric recognition, keyword biasing and turn detection. The model provides native transcripts and response text alongside audio.
GPT‑Live‑1 is not presented as a replacement for the entire agent stack. More demanding reasoning and actions can be delegated to an OpenAI or third-party backend text model, while developers retain control of models, tools and the agent framework. OpenAI’s Codex connection example omits connection setup and delegation handling, so it is not a complete implementation template.
In OpenAI’s own testing, GPT‑Live‑1 improved its Full Duplex Bench result by 30 percentage points over GPT‑Realtime‑2.1. Paired with GPT‑6 Astra at medium reasoning effort, the system ranked first in the Tau3 evaluation cited by OpenAI. These are vendor-reported results rather than independent findings.
OpenAI also cites an early evaluation by Speak in which interruptions during learners’ thinking pauses fell by almost 80% compared with earlier turn-based systems. The interactive demonstration included with the announcement is time-limited.
The architecture effectively separates conversational responsiveness from deeper processing: GPT‑Live‑1 manages the live voice exchange, while a developer-selected backend handles complex logic and tools. That provides flexibility, but it also requires additional components and costs beyond the voice layer.
GPT‑Live‑1 is available through the API at a stated price of $0.05 per minute for the front-end voice layer; the backend model and agent harness are separate. OpenAI has expanded support across accents, dialects and languages, but coverage remains incomplete, with more voices and languages promised over the following months. Custom voices are limited to eligible customers who contact sales and submit a request.
| Parameter | Conditions |
|---|---|
| Access | Through the API |
| Voice-layer price | $0.05 per minute |
| Complex reasoning | Separate backend text model |
| Custom voices | Available to eligible customers by request through sales |
| Languages and voices | Coverage is incomplete; expansion is planned over the following months |
Sources
Originally published by AI Feed: OpenAI Launches GPT‑Live‑1 for Full-Duplex Voice Apps.
Top comments (0)