GPT-Live-1 charges $0.05 per minute of voice, and its thinking is billed separately

What happened

• On September 10, OpenAI opened GPT-Live-1 in its API, the interface developers use to integrate its models into their own products. It’s the voice model ChatGPT already used: it listens and speaks at the same time, instead of waiting its turn.
• The price is $0.05 per minute for the voice layer. The model that reasons and carries out tasks behind it, whether GPT-6 Astra or one from another company, is billed separately.
• It includes support for phone calls, handling of interruptions and background noise, and an expanded catalog of voices with different accents.
• According to OpenAI, the language app Speak cut interruptions during its students’ pauses by almost 80%, and Yelp already uses it to answer calls for reservations and orders.

Why it matters

• A call center in Santiago, Bogotá or Lima can now calculate how much the conversational part of an agent costs: $3 per hour of conversation. With that number it can’t calculate the final bill, which depends on which model reasons behind it and how many times it’s queried on each call.
• The architectural change moves the cost elsewhere. Before, three pieces had to be coordinated: transcription, model and synthetic voice. Now voice is a single piece with a fixed price per minute, and the variable part sits with whatever model is connected to it.
• OpenAI says it will add voices and languages in the coming months. The announcement doesn’t specify how many voices speak Spanish or with what accent. In phone support in Chile, whether the agent sounds like Madrid or Mexico City isn’t a detail.

The number

$3 per hour. That’s what the voice of an agent that talks nonstop costs. Reasoning is paid separately, and Astra launched at $50 per million output tokens.

Context

Traditional voice agents chain together transcription, a text model and synthetic voice, and each handoff adds delay. OpenAI cites a healthcare company that eliminated 23,000 lines of code by replacing that chain with GPT-Live-1. The same day, the company introduced an agents API, and it offers Presence to large companies that want to deploy voice agents that resolve cases and hand off to people. The question of how much using AI really costs hasn’t changed. What changed is where the figure shows up.

What’s next

• New voices and languages in the coming months, with no dates.
• Custom voices only through the sales team, with an eligibility process.
• The agents API OpenAI introduced the same day remains in public beta.

Bottom line

On September 3, four AI assistants went down on the same day within a 90-minute window. An agent answering a company’s phone inherits that dependency, this time with someone waiting on the line.

Sources

• Build more natural voice experiences with GPT-Live-1 in the API — OpenAI, September 10, 2026
• GPT-Live guide in the API documentation — OpenAI Developers, accessed September 11, 2026

Edited by Rodrigo Cornejo. How we select and verify the facts, in who writes.