GPT-Live 1

GPT-Live 1 is a conversational voice model attributed by a third-party guide to OpenAI, released on September 10, 2026. Its defining feature is true full-duplex behavior, meaning it can listen and speak at the same time rather than taking rigid listen-then-respond turns. The guide frames the release as OpenAI's effort to make a voice assistant feel less mechanical, reducing the awkward pauses, premature responses to interruptions, and silent gaps that typically appear while a backend lookup runs.

Beyond the conversational layer, the guide describes an architecture that separates conversation rhythm from background work. The model is said to manage the timing and turn-taking of the spoken exchange, while a separate backend model or agent handles reasoning, search, calling approved tools, and returning a result. That division lets developers pair a natural-sounding voice front end with a more capable reasoning model behind the scenes, so a customer can keep talking while an order is being retrieved or a tool call is resolving. The practical fit is for voice agents and assistants where low-latency, interruption-friendly dialogue matters more than running the heavy reasoning inline.

Vercel AI Gatewayopenai/gpt-live-1gpt

Open provider page Provider docs

Quick Info

Provider: Vercel AI Gateway
Model key: openai/gpt-live-1
Release date: Sep 10, 2026
Last updated: Sep 10, 2026
Input modalities
Output modalities
Capabilities: Base catalog fields only

Limits

Output tokens: 0 tokens
Context window: 0 tokens

Latest news about GPT-Live 1

Vercel AI Gateway

CoverageBenchmark

OpenAI released GPT-Live-1 in its API on September 10, 2026, making the full-duplex voice model available to developers for the first time. The model can listen while speaking and delegates complex work to a separate backend model rather than handling all processing itself, a split that defines how the release should b The API spec listed in the Coursiv guide prices the voice session at $0.05 per minute, billed per second, with backend charges billed separately on whichever model and tools are paired. Inputs and outputs are audio and text only, with images and video unsupported; the model is restricted to live sessions and explicitly

Vercel AI Gateway

CoverageBenchmark

The llm-stats.com model page for GPT-Live-1 reports two benchmark scores attributed to OpenAI's official launch blog: TAU3-Bench at 0.86/1 using Tau3 (Voice) Intelligence Pass@1 averaged across airline, retail, and telecom domains, and Tau3 Banking at 0.32/1 using Tau Banking (Voice) Knowledge Pass@1 across 97 banking The page notes that GPT-Live-1 was paired with the GPT-6 Astra backend at medium reasoning effort in OpenAI's launch table, which is consistent with the split-architecture design of the voice model. It also situates GPT-Live-1 within leaderboards that include models such as Atria Dawn Preview, Ling 3.0 Flash, Grok 4.5,