OpenAI Releases GPT-Live – Next-Generation Voice Mode

OpenAI Releases GPT-Live – Next-Generation Voice Mode

(Actually, the release happened yesterday, but Grok-4.5 overshadowed it a bit)

So, what do we have:

— The model has been trained to listen and speak simultaneously. Even in Advanced Voice Mode, it was still a turn-taking exchange, but now the interaction feels like a live conversation (especially if you speak with the model in English). We tested it on simultaneous translation, and it sounds really good.

— The architecture has been replaced with Full-Duplex. This means that instead of the sequence "user -> processing -> response," the model now listens, updates, and thinks simultaneously. Specifically, every few seconds, it makes a decision: to continue listening, interrupt, call a tool, etc.

— If the request is complex, the system can redirect the question to GPT-5.5 (while naturally maintaining the conversation as the answer is being prepared).

The new mode (+ new voices) can be tried today. The free tier is rolling out GPT-Live-1 mini, while others should already have access to GPT-Live-1.

There is currently no video or screen demonstrations, and some languages sound with a noticeable accent, but developers promise to gradually fix all of this.