Skip to content
NewsProduct launch

OpenAI Opens GPT-Live-1 in the API at $0.05 per Minute with Full-Duplex Voice and Model Delegation

· by Pondero Newsdesk

The short version

The full-duplex voice model behind ChatGPT Voice is now available to all API users at $0.05 per minute, with support for delegating complex reasoning to GPT-6 Astra or third-party backends.

OpenAI Opens GPT-Live-1 in the API at $0.05 per Minute with Full-Duplex Voice and Model Delegation

The full-duplex voice model that has powered ChatGPT Voice since July 2026 went live in OpenAI's API on September 10, priced at $0.05 per minute for the front-end voice layer. Developers now have direct access to the same model that replaced the concatenated speech-to-text, language model, and text-to-speech pipeline in OpenAI's consumer product.

What

GPT-Live-1 handles audio input and output inside a single model rather than chaining three separate systems. That architecture removes the latency introduced at each handoff in traditional pipelines and lets the model catch interruptions in real time without pausing, because it continues to process incoming speech while it is speaking.

The delegation design is the most significant capability for production deployments. GPT-Live-1 manages conversation flow and can hand reasoning tasks or tool calls to a backend model, per Gigazine's coverage. Developers can pair it with GPT-6 Astra for demanding inference or with lighter models such as Luna when cost is the priority. Third-party backends are also supported, so the front-end voice layer is not locked to OpenAI's own reasoning stack.

Benchmarks cited by Unite.AI show GPT-Live-1 reaching 86.2% on Tau3 Voice Intelligence versus 45.7% for the predecessor GPT-Realtime-2.1. Turn-taking latency fell from 1.41 seconds to 0.798 seconds. Full Duplex Bench Interactivity rose from 45.4% to 80.10%.

Pricing covers the front-end voice layer only. At $0.05 per minute, a one-hour voice session costs $3.00 before backend model usage is added.

Why it matters

Until September 10, production-grade full-duplex voice required either an enterprise OpenAI agreement or a third-party provider such as ElevenLabs. The API opening changes the build calculus for teams working on phone agents, voice-first assistants, and accessibility tools.

The delegation architecture also shifts what is possible. A voice agent that routes reasoning to a capable backend without dropping conversation state is a different class of product from one that forces the developer to manage that handoff externally. EliseAI, an early design partner that builds AI communication tools for homes and healthcare, reported that GPT-Live-1 brought its phone agents to production viability, per Gigazine. Yelp, Speak, and Fin also reported improved call handling after early access, per Unite.AI.

For operators already running voice workloads, the $0.05 per minute front-end price is concrete enough to model against current spend. The number to check is total cost per conversation hour once the backend model is included, not just the voice-layer line.

What to watch next

Latency when GPT-Live-1 delegates to GPT-6 Astra under real workloads is the key variable. If the handoff adds perceptible delay in live calls, the benchmark advantage over GPT-Realtime-2.1 will narrow in practice. ElevenLabs Conversational AI is the closest published competitor on per-minute pricing; a direct cost comparison once backend charges are factored in will determine whether the OpenAI stack is competitive for high-volume deployments.

Sources