OpenAI’s voice AI no longer waits for your turn — it costs $0.05 a minute

Every voice AI on the market today tells you to wait before speaking. OpenAI’s GPT-Live-1 does not. The model, available through OpenAI’s API, listens and speaks simultaneously — handling interruptions, pauses, and overlapping speech the way a person does on a call.

The pricing is $0.05 per minute for the voice layer — $1.50 for a 30-minute conversation, or $50 per 1,000 conversation minutes. Developers pay separately for a reasoning backend: OpenAI’s own GPT-6 Astra or any compatible model that handles tool calls and business logic. The voice layer manages conversation mechanics; the reasoning layer decides what the AI actually says and does.

What that separation means in practice: a customer service bot can hear “wait, actually I want the other option” mid-sentence and respond to it, rather than plowing through a prepared answer. A language tutor can catch a student stumbling over a word without waiting for them to finish. A booking system can handle the kind of real-world call where people repeat themselves, change their minds, and speak across each other.

On the Full Duplex Bench — the evaluation designed specifically to test simultaneous-direction voice handling — GPT-Live-1 scores 30 percentage points higher than its predecessor, GPT-Realtime-2.1. When paired with GPT-6 Astra as the reasoning backend, it tops Tau3, the customer-service benchmark that tests complex, multi-step queries without conversation breaks.

The two-layer architecture reflects a deliberate design choice. Speech timing and reasoning are billed separately because they scale differently: a startup building a lightweight tutoring app pays less on both layers than an enterprise running thousands of simultaneous customer service calls on GPT-6 Astra. The floor is $0.05 a minute for voice, regardless of what runs behind it.

What GPT-Live-1 does not change is the quality of the answers. The voice layer handles when and how the AI speaks — not what it knows. A well-timed bad answer is still a bad answer. Developers building customer service applications will need to evaluate both layers separately, and the reasoning backend may add more to the bill than the voice layer itself. OpenAI has also not published data on how the full-duplex architecture performs in low-bandwidth environments or on noisy phone lines, where continuous audio input is harder to sustain.

GPT-Live-1 is now available in OpenAI’s API. The company has not announced a consumer-facing product built on the model, though it has indicated that future versions of ChatGPT’s voice mode may draw on the same underlying architecture. The ․0.05-per-minute rate applies at launch; OpenAI has not published a timeline for pricing changes or a list of third-party reasoning models approved for use with the voice layer.

Tags: artificial intelligence, OpenAI, voice AI, API, full-duplex audio, GPT-Live-1