OpenAI today announced that GPT Live 1 API is available to developers, bringing a single-model approach to full‑duplex voice interactions that listens and speaks simultaneously. By combining listening, speech generation and basic understanding in one model, OpenAI says GPT Live 1 API can respond to interruptions and acknowledgements in real time, simplify voice-agent architectures and reduce latency compared with traditional staged pipelines.
The company positions GPT Live 1 API as an alternative to the common three-stage design—speech-to-text, a reasoning model, and text-to-speech—that can introduce latency and brittle handoffs when users interrupt or change direction. With the new model, the front-end voice layer handles concurrent listening and speaking, and can delegate deeper reasoning or tool use to backend models or agent harnesses as needed. Developers still choose backend models based on throughput and reasoning complexity, allowing the voice layer to focus on natural interaction while heavier tasks run elsewhere.
OpenAI’s documentation and examples show typical pairings that balance speed and capability. For tasks that require higher throughput, developers can pair GPT Live 1 API with models such as Luna for scheduling and order updates. For more complex reasoning or tool-driven tasks, the company cites pairings with Astra or Terra. The model natively returns ASR transcripts alongside response text, supports alphanumeric understanding and keyword biasing, and offers turn detection for workflows that require explicit turn boundaries.
On performance, OpenAI reports measurable gains in full‑duplex voice metrics. In internal evaluations, GPT Live 1 API improved Full Duplex Bench performance by 30 percentage points compared with GPT‑Realtime‑2.1, with reductions in turn-taking latency and improvements in interactive behavior. When paired with GPT‑6 Astra at medium reasoning effort, OpenAI says GPT Live 1 placed first on Tau3, a benchmark the company describes as measuring frontier voice-agent intelligence on end-to-end tasks.
OpenAI also cites an early customer result from language-learning company Speak, which found that GPT Live 1 API reduced interruptions by nearly 80% compared with previous turn‑based systems. In Speak’s use case, the full‑duplex tutor gave learners more time to think before responding, improving the interaction flow in a learning context.
Real‑world audio robustness is another focus. GPT Live 1 API is designed to manage background noise and silence without narrating every step, and to support side conversations, backchannels and interruptions. These capabilities are intended to maintain conversational flow in noisy environments—such as coffee shops or city streets—and to improve reliability across extended sessions.
OpenAI is also expanding real‑time voice options alongside the launch, offering more accents, dialects and languages so developers can choose how assistants sound. The model supports telephony, enabling full‑duplex voice agents for phone-based tasks such as reservations or customer support, according to the company.
On pricing and availability, GPT Live 1 API is available in the API now at a front‑end voice layer price of $0.05 per minute. Developers can combine the front-end model with backend models and agent harnesses that match a task’s reasoning depth, latency and cost requirements. For custom voice access and enterprise integrations, OpenAI directs customers to contact sales or account directors for offerings such as OpenAI Presence, which uses the model to power real‑time voice workflows for enterprises.
To help developers and evaluators test the model’s capabilities, OpenAI provides a time‑limited demo where users can start a session, speak naturally, interrupt responses and try the model in noisy settings. Developer documentation outlines APIs and guides for building live voice experiences and for delegating tasks from the voice layer to backend models.
By consolidating listening and speaking into a single front‑end model while enabling backend delegation, GPT Live 1 API aims to simplify voice-agent stacks and make conversations feel more natural. Early benchmarks and customer feedback suggest meaningful gains in latency, interruption handling and interactive behavior, while telephony and expanded voice options broaden potential deployment scenarios. As developers experiment with pairings and production integrations, the new offering will be one to watch for voice-first applications.
Source: Read the original source
Leave a Reply