GPT-Live-1: Full-Duplex Voice Arrives on Taliiq
OpenAI's GPT-Live-1 listens and speaks at the same time. Here's what full-duplex actually changes for a Taliiq voice agent, and how it's wired up.
What "full-duplex" actually means for your callers
Most voice models, however fast, still work in turns under the hood. The model listens, then speaks, then listens again. GPT-Live-1, which OpenAI shipped on September 10, 2026, works differently. It processes incoming and outgoing audio at the same time, so a caller can talk over it, laugh mid-sentence, or change their mind without the awkward silence-then-restart most voice bots produce.
The numbers back it up. OpenAI puts GPT-Live-1 thirty percentage points ahead of GPT-Realtime-2.1 on Full Duplex Bench, and turn-taking latency drops from 1.4 seconds to 0.8. In practice that means fewer "sorry, go ahead" moments and more calls that feel like talking to someone who's actually paying attention.
Where the reasoning actually happens
GPT-Live-1 doesn't try to be the reasoning engine too. It handles listening, speaking, and deciding when to hand a turn off, and OpenAI pairs it with a backend model on their own infrastructure to do the actual thinking and tool calls. That split is entirely inside OpenAI's stack, not a setting you configure per call. On Taliiq, picking GPT-Live-1 as your realtime model means it runs the full voice turn end to end. The Brain LLM tab you'd normally use to pick a reasoning model only applies to web widget chat and Taliiq's cascaded STT to LLM to TTS pipeline, not to realtime voice. The same split still paid off for one early adopter, a healthcare company, who cut 23,000 lines of cascaded voice glue code (interruption handling, pipeline orchestration, audio sync) down to a single voice layer call.
Live on Taliiq now
GPT-Live-1 is available today as a model choice under the Realtime Model tab, next to the existing gpt-realtime family and xAI's Grok Voice. It comes with a broader voice library than the original Realtime API, gives you transcripts and turn-detection out of the box even though it isn't strictly turn-based.
If your agent deals with background noise, side conversations, or callers who interrupt themselves mid-thought (a restaurant taking orders, a clinic booking appointments), it's worth putting GPT-Live-1 against your current realtime setup on the same call scenarios you already test.
Put it on a real call today
Switch the model in the Voice tab and hear the difference yourself. Setup takes minutes, no card required to start.