FULL STORY

OpenAI Launches GPT-Live-1 Voice Model on API

OpenAI brought ChatGPT's natural full-duplex voice experience to developers via the GPT-Live-1 API. Follow-up analysis pegged its $0.05-per-minute pricing at roughly $28 per million tokens.

2026-09-11 ~ 2026-09-12 · 2 episodes · 39 posts

Episode 1 · OpenAI brings full-duplex speech model GPT-Live-1 to its API (2026-09-11, 37 posts)

On September 11, OpenAI officially announced that the speech model GPT-Live-1 is now live on the API, opening up ChatGPT's natural full-duplex conversational experience to developers: voice agents can listen while speaking, support user interruptions and natural back-and-forth, and developers can freely pair the voice interaction frontend with a text model and harness of their choosing. The company says it achieves unprecedented turn-taking naturalness among the most advanced speech models.

Confirmed

  • Interruption support: users can add details or change topics mid-utterance without waiting for the model to finish its turn.
  • Noise robustness: the model can distinguish speech from background noise; an official demo showed a conversation continuing in a noisy café without breaking off.
  • Customizability: developers can use instructions to set a voice agent's tone, pacing, and expressiveness; the model mirrors the emotional tone of the speaker and adapts to their speaking speed, and also supports setting language and response length.
  • Architectural flexibility: the speech model can act as an interaction frontend, delegating tasks to a text model of the developer's choice.

Why it matters

  • This is the first time OpenAI has exposed ChatGPT's real-time full-duplex voice capability as an API, letting developers build near-human conversational experiences into their own voice agent products with more control over how agents speak and behave.
  • Decoupling the voice frontend from the text reasoning backend means developers aren't locked into a single model stack, offering far greater flexibility for building voice agents.

17 more related posts →

Episode 2 · GPT-Live costs ~$28/M tokens, hands off reasoning to backend models (2026-09-12, 2 posts)

GPT-Live charges $0.05/minute, which works out to roughly $28/M tokens assuming 12.5Hz audio tokens. Unlike GPT-Realtime, it only handles conversation while delegating reasoning and tool calls to backend models.