How OpenAI built a realtime voice AI system in six months

OpenAI says it built GPTLive, a thirdgeneration voice system for ChatGPT, to make voice conversations feel more immediate and natural. The company described a fullduplex model that can listen and speak at the same time, while handling deeper reasoning and tool use on a separate asynchronous path. The engineering work focused on lowering latency across the entire stack. OpenAI said it reworked model inference, context management, and media transport, and moved media handling to a dedicated fast path so application logic would not slow down live speech. The post also outlines production techniques used to keep long voice sessions stable, including seamless model handoffs, context compaction, and faster session setup with WARP and Instant Connect. OpenAI said the architecture is already supporting ChatGPT Voice and will also underpin an upcoming GPTLive API.