OpenAI's engineering team details how they built GPT-Live, a third-generation voice AI system, in six months. The system replaces the traditional turn-detector architecture with a full-duplex voice model that can listen and speak simultaneously. Key engineering decisions include: separating the media fast path from application logic, rewriting the media frontend in Go (replacing Python asyncio) for smoother frame delivery, using WebRTC as the transport layer, and developing WARP (WebRTC Abridged Roundtrip Protocol) to reduce session startup from six network round trips to one. The post also covers stateful inference with seamless model instance handoffs, dynamic context compaction without media interruption, asynchronous delegation to frontier models like GPT-5.5, and an 'Instant Connect' mechanism for near-instant session establishment. Production validation was done via shadow testing on real ChatGPT Voice traffic, which revealed capacity, geography, and observability challenges not visible in synthetic load tests.
