GPT-Live-1: OpenAI's voice model can listen and talk at the same time

What's new

On September 10, 2026, OpenAI launched GPT-Live-1 in its API, the technical access point developers use to build OpenAI's models into their own applications. It is a so-called full-duplex voice model, meaning it can listen and speak at the same time, just like in an ordinary conversation between two people.

The model supports telephony for phone agents, including outbound calls. It can detect when it's your turn to speak, understand strings of letters and numbers, and filter out background noise. It also offers transcriptions and the option to weight specific keywords. At launch there are 12 voices to choose from. Those who want their own custom voice must contact OpenAI's sales team.

What's smart

Traditionally, a voice assistant is built as a chain: speech is first converted to text, then a language model does the thinking, and finally the text is turned back into speech. GPT-Live-1 processes incoming and outgoing audio together in a single model. According to Pasquale Pillitteri's review, this should provide more natural handling of interruptions and of small acknowledging sounds like "mm" and "yeah."

The model is designed as a voice layer on top of the models that handle the actual thinking in the background. The articles mention GPT-6 Astra and Luna. On the other hand, GPT-Live-1 only accepts audio and text input. Image and video input are not supported, even though image input was a feature of its predecessor. The model's knowledge cutoff is July 31, 2025.

Cheaper or better?

The price is $0.05 per minute of session time for the voice layer. Backend models and tools are billed separately, and according to the sources, the phone provider's fees come on top. The predecessor GPT-Realtime-2.1 was billed per audio token (chunks of text): $32 per million input tokens and $64 per million output tokens. There is no direct comparison of what a conversation costs with the two models. There is no free tier. Tier 1 allows 25 concurrent sessions, and Tier 5 allows 500.

On quality, the figures are OpenAI's own measurements and have not been independently verified. According to DataNorth, the model scores 30 points higher than GPT-Realtime-2.1 on Full Duplex Bench, but the absolute scores were not disclosed. Against GPT-Realtime-2, Tau3 Voice Intelligence is 86.2 percent versus 45.7 percent, but that is OpenAI's own measurement for GPT-Live-1 together with GPT-6 Astra (medium reasoning), and it applies neither to the voice model alone nor to its use with cheaper models. Turn latency, meaning the wait before the response, is 0.798 seconds versus 1.63 seconds, and Conversational Dynamics is 97.3 percent. DataNorth notes, however, that latency figures have not been published.

Competitors mentioned include Microsoft's MAI Realtime, which is also full-duplex, xAI's Grok Voice and Block's Dessa, which is trained solely on speech data. The sources do not provide a direct comparison with them.

What it's good for

GPT-Live-1 is aimed at businesses and developers who want to build phone agents and voice assistants, for example for customer service or outbound calls. Here, the ability to handle interruptions and noise makes a difference, because the conversation needs to feel less stiff.

According to the sources, the model is available globally right away. The sources do not say whether Danish is supported or whether data is processed in the EU. Developers must take into account laws on call recording and personal data in the countries where they operate. For ordinary users, this is a technology they will encounter indirectly when companies use it in their phone systems.