Voice AI’s Full-Duplex Shift: Smarter, but Still Scripted

OpenAI’s new voice model is not a monolithic intelligence. It splits responsibilities: a lightweight duplex model handles real-time conversational flow, while GPT-5.5 silently takes over for search, fact-checking, and reasoning. This delegation pattern lets the voice layer stay fast and natural, while still delivering heavyweight answers.

For teams building voice interfaces, the turn-based assumption dissolves. The model listens while speaking, manages interruptions, corrects grammar mid-sentence, and pauses to gather context before translating. UX design must now treat voice AI as a proactive participant, not a command responder.

The demos are impressive but entirely scripted. No benchmarks, latency numbers, or failure modes are shown. Real‑world robustness—handling accents, noise, overlapping corrections—is unproven. Safety claims are vague, and always‑listening mode raises privacy and governance questions engineering leaders cannot ignore.

Cost and infrastructure tradeoffs are also hidden. Delegation adds a hidden server‑side hop; if the text model is bottlenecked, the voice experience degrades. Teams adopting this pattern must weigh real‑time performance against backend complexity. The signal is clear: voice AI is becoming an orchestration challenge, not just an ML one.