Topics›Industry and Market Signal›How OpenAI Built GPT-Live: 'Separating Talking From Thinking' in a Full-Duplex Voice Model
★MediumAct: LowFull content
How OpenAI Built GPT-Live: 'Separating Talking From Thinking' in a Full-Duplex Voice Model
ByteByteGo·September 22, 2026
New
Summary
A small fast voice model handles conversation while a frontier model (GPT-5.5) is delegated complex reasoning/tool calls asynchronously so conversation never stalls. Evaluation for full-duplex systems differs fundamentally from turn-based systems: no discrete turns, so teams score endpointing/barge-in behavior and monitor p999 (not p95/p99) latency.
Why this signal
Interesting architecture pattern (delegate reasoning asynchronously to keep an interactive loop responsive) that could generalize beyond voice to any latency-sensitive agent UI, though not a direct Thomas Engineering application today.
Original email
ByteByteGo · 9/22/2026
How OpenAI Built GPT-Live