TL;DR: OpenAI released GPT-Live-1, a full-duplex speech-to-speech model that delegates reasoning and tool use to a configurable backend text model, and it debuted at #1 on Artificial Analysis's Speech to Speech Index.
Summary: OpenAI's GPT-Live-1 is a new full-duplex speech-to-speech model that keeps the conversation going while handing reasoning and tool use off to a separately configured backend text model; developers stream audio in and receive speech back via the API. On Artificial Analysis's Speech to Speech Index it scored 81.5 with the Astra backend at medium reasoning effort (rank #1) and 80.1 with Sol at low reasoning effort (rank #3), just ahead of and behind Grok Voice Think Fast 2.0 High at 81.3. It also took the top two spots on the Tau Voice agentic benchmark at 67.9% and 59.3%, though it trailed on audio reasoning (90.1% on Big Bench Audio vs 97.2% for Grok Voice Think Fast 2.0 High).
Why it matters: Decoupling the realtime voice layer from a swappable reasoning backend changes how builders architect voice agents — you can upgrade or route the text model without retraining the speech stack. Worth testing against Gemini 3.1 Flash Live and Grok Voice Think Fast on task success rates, where GPT-Live-1's lead is narrower than its index ranking suggests.
Source: x_com