OpenAI Releases GPT-Live-1 Full-Duplex Voice Model in API

GPT-Live-1 is a unified full-duplex voice model from OpenAI that combines listening and speaking functions into one system without separate ASR, TTS and LLM stitching.

OpenAI introduced GPT-Live including the GPT-Live-1 and GPT-Live-1 mini models on July 8, 2026. The models power ChatGPT Voice and became available through the public API on September 10, 2026 with full documentation, evaluations and pricing details released at the same time.

What background led to the GPT-Live-1 release?

Voice systems before GPT-Live-1 relied on cascaded pipelines that connected separate automatic speech recognition, text-to-speech and language model components. These stitched systems introduced latency and struggled with natural overlaps such as interruptions or backchannels during speech.

The GPT-Live System Card states that GPT-Live-1 and GPT-Live-1 mini represent a new generation of voice models designed to make conversations with AI feel more natural and intelligent through continuous listening and responding instead of waiting for defined turns.

What new capabilities does GPT-Live-1 deliver?

The model listens while speaking without requiring explicit turn boundaries. It handles human-like interruptions, provides backchannels such as affirmations, and maintains conversational flow even when delegating deeper reasoning or search tasks to backend models.

GPT-Live-1 delegates complex agentic tasks to models such as GPT-5.5 or GPT-6 Astra during API evaluations while the front-end voice layer continues to manage real-time audio exchange. This separation keeps the conversation natural while offloading heavy computation.

How does the unified architecture function?

A single model processes incoming audio and generates outgoing audio simultaneously. The full-duplex design removes the need for explicit handoffs between recognition, reasoning and synthesis stages that previously created artificial pauses.

The architecture supports continuous response generation so the model can react to user input mid-sentence when appropriate. This capability stems from research advances noted in the system card that enable listening and responding at the same time.

What performance gains appear in evaluations?

OpenAI reports large gains in turn-taking latency and interactive behavior on internal benchmarks. The model shows measurable progress in handling overlapping speech and maintaining context across extended exchanges.

What do partners report about real-world use?

Education and healthcare developers have integrated the model into production workflows. The changes allow teams to focus on higher-level experience improvements rather than low-level pipeline management.

A good language tutor knows when to give learners space and when to step in, and GPT‑Live‑1 brings that naturalness to Speak’s Live Tutor Lessons—in our early evaluations, it cut interruptions during thinking pauses by almost 80% compared with previous turn-based systems.Andrew Hsu, Co-founder & CTO, Speak

A separate integration report noted that the unified model removed thousands of lines of custom code required for component coordination in earlier cascaded builds.

Compared to our cascaded build, GPT‑Live‑1 simplified our code base by 80% and removed 23K lines of code. This enabled natural, real-time patient conversations & freed our team to improve the experience from booking an appointment to navigating care.Tony Stoyanov, Co-Founder & CTO

What are the availability and pricing details?

GPT-Live-1 is the default voice model for paid ChatGPT users on Go, Plus and Pro plans. Free users receive GPT-Live-1 mini. Both models are accessible through the OpenAI API with public documentation and evals.

API voice sessions carry a front-end voice layer price of $0.05 per minute billed per second. Backend model usage for reasoning or tool calls is billed separately according to the respective model rates.

What market implications follow from the release?

Developers building voice-enabled apps and business workflows gain a simpler integration path. The reduction in custom code allows faster iteration on features such as appointment booking or care navigation in healthcare settings.

Education platforms can now deliver tutor-style interactions that respect learner thinking time without premature interruptions. This change supports broader adoption in conversational tutoring and customer service scenarios.

1. Review the public API documentation and evals released on September 10, 2026.
2. Configure voice session parameters with $0.05 per minute front-end pricing.
3. Delegate complex tasks to GPT-5.5 or GPT-6 Astra as needed for the use case.
4. Test turn-taking and interruption handling on Full Duplex Bench metrics.
5. Deploy the model in production for ChatGPT Voice or custom applications.

How does GPT-Live-1 compare to prior models?

Comparison of OpenAI voice models released in 2026

Model	Architecture	Full Duplex Bench Gain	Primary Availability
GPT-Realtime-2.1	Cascaded ASR-TTS-LLM	Baseline	Prior generation
GPT-Live-1	Unified full-duplex	30 percentage points	September 10, 2026 API
GPT-Live-1 mini	Unified full-duplex	Optimized variant	Free ChatGPT users

What developments are expected next?

Further iterations may extend the full-duplex approach to additional languages and multi-modal inputs. Continued research advances could enhance delegation mechanisms for more sophisticated agent behaviors while preserving conversational naturalness.

The system card indicates that these models result from ongoing research into continuous audio processing. Future releases are likely to build on the same unified architecture for expanded enterprise and consumer applications.