OpenAI releases GPT-Live-1 to the API with full-duplex voice and modular delegation (2026)

OpenAI released GPT-Live-1 to its API on September 10, 2026, making the full-duplex voice model behind ChatGPT Voice available for developer applications. The model processes incoming and outgoing audio simultaneously within a single architecture, bypassing the serial delays of traditional speech-to-text, reasoning, and text-to-speech pipelines. OpenAI is pricing the front-end voice interaction layer at $0.05 per minute, while allowing developers to route background reasoning to arbitrary text models. 1 2 3

What launched

Signal	Confirmed detail	Action window
Full-duplex audio processing	A single neural architecture processes input audio while generating output speech, replacing cascaded pipelines that require separate transcription, inference, and synthesis stages. 2	Benchmark latency and interruption handling against existing multi-stage voice infrastructure.
Decoupled reasoning delegation	The voice layer delegates complex tool execution, database queries, and deep reasoning to backend models such as GPT-6 Astra, Luna, or Codex while maintaining continuous conversational presence. 2 3	Configure lighter models for simple interactions and route multi-step tasks to frontier reasoning tiers.
Continuous interaction decisions	The model assesses conversational context many times per second to decide whether to speak, listen, insert vocal acknowledgments, pause, or handle an interruption. 2	Test mid-sentence user interruptions and ambient noise tolerance in noisy operational environments.
Telephony and audio protocols	The API supports WebRTC, WebSockets, and SIP connections for direct deployment across mobile applications and phone networks. 2 3	Review network transport architecture before staging customer-facing call routing.
Usage-based voice pricing	The voice interaction layer costs $0.05 per minute. Backend model tokens and tool invocations incur standard API billing rates, with custom voice access restricted to enterprise sales agreements. 2	Model combined per-minute voice fees and per-token backend compute costs for high-volume call centers.

Conversational benchmarks and implementation constraints

On OpenAI's Full Duplex Bench evaluation, GPT-Live-1 achieved a 30 percentage point gain over GPT-Realtime-2.1 in turn-taking latency, natural backchanneling, and interruption recovery. When paired with GPT-6 Astra at medium reasoning effort, the system placed first on Tau3, an evaluation evaluating spoken customer-support workflows across airline, retail, and telecommunication tasks. 2

In commercial deployments, Speak reported that the model reduced false interruptions during learner thinking pauses by roughly 80% compared with prior silence-based turn-detection systems. Yelp and Hatch integrated the model to automate restaurant telephone reservations, while Cognition adopted it to provide real-time voice steering for Devin engineering sessions. 2

The initial release contains clear functional boundaries. Video inputs and live screen sharing remain unsupported. OpenAI provides no free API tier for the voice layer, and weights remain proprietary and cloud-hosted. 2 3

Why it matters

GPT-Live-1 formalizes a separation between conversational delivery and analytical computation. Chained voice pipelines suffered from cumulative latency and brittle turn boundaries that triggered whenever a speaker paused to collect their thoughts. By handling voice timing as a continuous audio-native process while delegating cognition to external models, OpenAI provides developers with a persistent conversational interface that remains stable even as underlying reasoning models receive updates. Teams building customer support agents, language tutors, and interactive assistants can adopt the $0.05-per-minute voice API to eliminate custom turn-taking code, while monitoring backend token consumption as complex queries execute in parallel.