GPT-Live-1 API Review 2026: Full-Duplex Voice Agents, Pricing and Features

GPT-Live-1 is now available through the OpenAI API, bringing ChatGPT’s full-duplex voice system to developers. Released on September 10, 2026, the model can listen and speak at the same time, respond to interruptions, handle silence and background noise, and delegate deeper reasoning or tool use to another model.

For companies building phone agents, tutors, assistants, games, and customer-service systems, the important change is architectural: GPT-Live-1 can replace part of the traditional speech-to-text, language-model, and text-to-speech chain with one continuous conversational layer.

What full-duplex voice means

Most voice bots wait for the user to finish, transcribe the audio, generate text, and synthesize a spoken response. GPT-Live-1 processes incoming and outgoing audio together, allowing it to decide continuously whether to listen, speak, pause, acknowledge, interrupt, or call a tool.

This can make conversations feel less like alternating recorded messages. Users can correct the agent mid-sentence, hesitate, laugh, briefly speak to someone nearby, or add detail without restarting the interaction.

Key GPT-Live-1 API features

Improved interruption handling

The model reasons over both directions of audio in one system. OpenAI reports that language-learning company Speak saw almost 80% fewer unwanted interruptions than with previous turn-based systems in an early evaluation.

Reasoning and tool delegation

GPT-Live-1 can pass a difficult question or action to a backend model such as GPT-6 Astra while keeping the voice interaction alive. Developers choose the reasoning model, tools, and agent harness, allowing a lightweight backend for scheduling and a stronger model for complicated support issues.

Tone, pace, and conversational style

System instructions can shape how the agent speaks, including tone, pacing, directness, and style. This is important for branded support, coaching, education, hospitality, and accessibility applications.

Noise, silence, and long sessions

The model is designed to distinguish pauses and background conversation from an actual end of turn. OpenAI also says it retains context and conversational quality more reliably during extended sessions.

Telephony

GPT-Live-1 supports phone-based voice agents for restaurant reservations, order handling, sales qualification, and customer service. Applications remain responsible for consent, recording disclosures, identity checks, and human escalation.

Does it replace speech-to-text and text-to-speech?

For some applications, GPT-Live-1 can simplify the voice layer because it understands and generates audio continuously. It also provides ASR transcripts and response text. More controlled systems may still combine it with separate components for compliance transcripts, specialized voices, or deterministic processing.

Performance claims

OpenAI reports a 30-percentage-point improvement over GPT-Realtime-2.1 on Full Duplex Bench. When paired with GPT-6 Astra at medium effort, it also ranks first on OpenAI’s cited Tau3 evaluation for end-to-end voice-agent tasks. These are vendor-reported results, so teams should test their own accents, languages, noise conditions, and workflows.

New voices and language coverage

The API launch expands the available selection across accents, dialects, languages, and speaking styles. Named options include Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, and Cinder. OpenAI says coverage will continue expanding.

GPT-Live-1 pricing

The front-end voice layer is available at $0.05 per minute. Backend model tokens, tools, telephony, storage, and agent infrastructure can add to the total cost. Estimate complete call cost rather than comparing only the headline per-minute rate.

Best use cases

• Customer-support and reservation phone agents
• Language tutors and conversational learning
• Hands-free workplace assistants
• Interactive characters, games, and accessibility tools
• Voice interfaces backed by Codex or GPT-6 Astra

Limitations and safety

Voice systems can mishear names, numbers, addresses, consent, and emotionally sensitive language. High-impact actions should require explicit confirmation, and users should know when they are speaking with AI. Applications also need interruption tests, fallback behavior, human handoff, abuse controls, and privacy policies for stored audio and transcripts.

Verdict

GPT-Live-1 is a meaningful step toward voice agents that behave like participants in a conversation instead of a sequence of audio requests. Its strongest advantage is the combination of natural turn-taking with delegation: the voice layer can stay responsive while another model or tool handles deeper work.

Explore the official GPT-Live-1 API guide

Source note: Based on OpenAI’s September 10, 2026 release announcement and API documentation. Features, prices, and voice availability may change.