GPT-Live-1 API: Build Voice Agents for $0.05 a Minute

What GPT-Live-1 Changes for Voice Developers

OpenAI has opened GPT-Live-1 to developers. The full-duplex voice model that has powered ChatGPT Voice since July reached general availability in the OpenAI API on September 10, and it brings a different architecture from the voice models most developers are using today. Instead of taking turns (listen, stop, think, speak), GPT-Live-1 listens and talks at the same time, and it hands the heavy reasoning to a separate backend model while it keeps the conversation moving.

• Price: $0.05 per minute of session time, billed per second, with backend model and tool usage billed separately at standard API rates
• Endpoint: the new Live endpoint (v1/live/sessions) only, not Chat Completions, Responses or Realtime
• Benchmark (OpenAI's figure): 86.2% on the Tau3 voice-agent benchmark, versus 45.7% for GPT-Realtime-2.1 and 42.4% for GPT-Realtime-2
• Concurrency: 25 simultaneous sessions at usage Tier 1, rising to 500 at Tier 5

How Does a Full-Duplex Voice Model Work?

Most voice agents built in the last two years are turn-based under the hood. The system waits for the caller to stop, transcribes, reasons, then speaks, and anything the caller says in the meantime is either lost or treated as an interruption that resets the exchange. A full-duplex model keeps both channels open. GPT-Live-1 can be cut off mid-sentence, pause naturally, and call tools during the conversation without the awkward dead air that makes phone bots feel like phone bots.

The more interesting design choice is the split between talking and thinking. When a request needs real work, such as checking availability, searching records or planning a multi-step booking, GPT-Live-1 delegates it to a backend model or agent (GPT-6 Astra, or a cheaper model for simpler jobs) and keeps the caller engaged while that work runs. The voice layer is tuned for timing and conversational behavior; the backend is chosen for the task and the budget. A support line might pair it with a small, fast model, while a research assistant might pair it with a frontier reasoner.

According to OpenAI's model documentation, inputs and outputs are audio and text (no images or video). Streaming and function calling are supported; structured outputs and fine-tuning are not.

What the Early Numbers Show

The benchmark claims are large, and they are OpenAI's own. Beyond the Tau3 result, the company reports an improvement of roughly 30 over GPT-Realtime-2.1 on Full Duplex Bench. Independent evaluations have not been published yet, so treat these as a starting point rather than a verdict.

The customer reports are more concrete. Language-learning app Speak reports close to 80% fewer interruptions than its previous system, according to The Register, and Yelp has deployed the model behind Yelp Host, its restaurant reservation assistant. Custom voices are available through OpenAI's sales team.

On the same day, OpenAI's API changelog also listed its Agents API in public beta. That hosted runtime for long-running agents is a natural candidate for the backend half of a GPT-Live-1 deployment.

Is GPT-Live-1 Worth Adopting Now?

The cost model is the first thing to work through. At $0.05 a minute, the voice layer alone runs $3 per hour of conversation before backend tokens and tools, which is easy to budget for customer-facing lines where a human call costs far more. The payoff is architectural: fewer stalls, natural interruptions, and the freedom to swap the backend model without touching the voice experience.

The practical caveat is migration. Because GPT-Live-1 runs only on the new Live endpoint, moving from GPT-Realtime means adopting a new session API rather than changing one model string. For a sense of how fast this stack has moved, compare it with GPT-Live's July debut inside ChatGPT and the earlier gpt-realtime-2.1 speed and price update. More model launches are in our AI news coverage.

Sources: OpenAI — September 10, 2026; OpenAI API docs: GPT-Live-1 — September 2026; OpenAI API changelog — September 10, 2026; The Register — September 10, 2026.