OpenAI's GPT-Live Removes Turn Detector for Real-Time Voice

The core problem with earlier voice systems was architectural: they inherited the turn-based nature of text models, forcing a separate, small model to pick the exact moment to respond. Guessing wrong meant interrupting the user or creating an awkward pause. GPT-Live addresses this not by improving the detector, but by removing it entirely, placing a full-duplex model in control of a continuous audio stream.

This shift required a substantial re-engineering effort, moving from a request-response inference model to a stateful, streaming one. The team wrote the media frontend in Go to improve frame delivery, developed a seamless handoff mechanism for long-running calls and context compaction, and created WARP and Instant Connect to reduce session startup time. The silent production test revealed that capacity is not just about GPU throughput but about sustaining concurrent sessions while keeping every audio frame on schedule, and that geography and observability are first-order concerns.

The architecture is positioned as a broader platform, powering ChatGPT Voice's expansion into agentic coordination and set to underpin the upcoming GPT-Live API. How well the new system handles the unpredictable nature of real-world network conditions and long, complex conversations will be the real test, and the API's performance will show how this design translates for external developers.