On 10 September 2026, OpenAI made its full-duplex voice model GPT-Live-1 available in the API. The same model that powers ChatGPT’s voice mode is now open to any developer building custom voice agents.
Listening and speaking at once
Unlike traditional turn-based voice APIs, GPT-Live-1 keeps processing incoming audio while it is speaking. Callers can interrupt mid-sentence without breaking the conversation. According to OpenAI, the model ships native transcripts, turn detection, better handling of background noise and telephony support for phone-based deployments. Twelve voices are available at launch.
Pricing and trade-offs
OpenAI prices access at $0.05 per minute, but that covers only the voice layer. Whatever backend model does the reasoning is billed on top. One caveat: the predecessor GPT-Realtime-2.1 accepted image input, a capability GPT-Live-1 drops according to the reports available so far.
Benchmark is a vendor claim
OpenAI says GPT-Live-1 improves performance on its in-house “Full Duplex Bench” by 30 percentage points over GPT-Realtime-2.1. This is a vendor figure; independent numbers were not available at launch. Whether the model lives up to the demo in production will only become clear once third-party tests land.
Sources: OpenAI announcement and reporting by DataNorth (as of 10 Sept 2026).
