GPT-Live-1 Comes to the API: What It Means for Voice Agents

*This research-based analysis covers OpenAI's September 10, 2026 API release. Performance figures and early-customer results are OpenAI's reported evaluations unless otherwise stated. Pricing and capabilities can change.*

The short answer

GPT-Live-1 is now available through the OpenAI API as a full-duplex voice layer priced at $0.05 per minute. It can listen while speaking, handle interruptions and brief acknowledgments, produce transcripts and response text, support telephony, and delegate reasoning or tool calls to a separate backend model such as GPT-6 Astra.

The release targets a persistent weakness in voice agents: chained speech-to-text, language-model, and text-to-speech systems often lose timing and conversational context at each handoff. A single front-end model may make turn-taking feel more natural. It does not eliminate backend-model charges, telephony, tools, infrastructure, monitoring, or human escalation.

What is full-duplex voice?

Full duplex means the system can process incoming speech while producing outgoing audio instead of forcing rigid alternating turns. That enables a caller to interrupt, correct a detail, laugh, hesitate, or give a short acknowledgment without the entire interaction resetting.

OpenAI says GPT-Live-1 reasons across incoming and outgoing audio together, manages silence and background noise, and retains quality across longer sessions. It also exposes transcripts, keyword biasing, and turn detection for applications that still need explicit boundaries.

How much does GPT-Live-1 cost?

OpenAI lists $0.05 per minute for the front-end voice layer. A ten-minute conversation therefore starts at $0.50 for that layer. The full cost can also include the delegated text or reasoning model, agent harness, tool calls, phone carrier, storage, observability, and review.

Model cost per resolved conversation, not merely per minute. Track containment rate, transfers, repeated questions, tool errors, user abandonment, review labor, and any credits or remediation caused by a bad interaction.

What can developers build?

The launch describes browser connections through WebRTC, server integrations through WebSockets, and phone workflows through telephony. Likely uses include support, reservations, tutoring, intake, internal help desks, and voice access to a backend agent. OpenAI also offers a sales-led enterprise layer called Presence.

Delegation creates a useful architecture: GPT-Live-1 manages conversational timing while a backend model or tool handles deeper work. It also creates a handoff that must be tested for delay, duplicated actions, stale context, and what the voice layer says while the backend is still working.

What should teams test before launch?

Evaluate real accents, dialects, speech differences, code-switching, names, account numbers, noisy rooms, weak connections, long silences, emotional callers, interruptions, and simultaneous speech. Test hang-ups and reconnects, tool timeouts, transfers, identity verification, consent, recording notices, emergency language, and explicit confirmation before consequential actions.

Users should know they are speaking with AI. Custom voices require appropriate rights and consent, and organizations need policies for recordings, transcripts, retention, sensitive information, and access. A natural voice raises the risk of misplaced trust; disclosure and escalation should become clearer as realism improves.

The verdict

GPT-Live-1 makes a credible architectural change by separating natural full-duplex conversation from the backend intelligence and tools a workflow needs. The published $0.05-per-minute rate makes early cost modeling straightforward, but it is only the front layer.

Pilot one narrow workflow where interruptions matter, compare it with the current system, and require safe escalation. The goal is not the most human-sounding demo; it is a correctly resolved conversation with informed users, controlled actions, and a complete evidence trail.