The new full duplex speech model from OpenAI delegates reasoning and tool use to a backend text model while streaming audio via API. Using Astra at medium reasoning effort, the GPT-Live-1 model achieved 81.5 on the Artificial Analysis Speech to Speech Index, placing it ahead of Grok Voice Think Fast 2.0 High at 81.3. Costs for the model are $4.47 per hour for the Sol backend and $5.83 per hour for Astra, inclusive of session charges and token usage.

Gemini 3.1 Flash Live Minimal leads conversational preference with 1,096 Elo, while GPT-Live-1 Sol ranks third at 1,053. GPT-Live-1 holds the top two spots on the Tau Voice agentic benchmark at 67.9% and 59.3%, though these figures are based on a single trial. Average time to first audio for the model is 1.24 to 1.34 seconds, compared with 0.70 seconds for Grok.