OpenAI has officially opened API access to GPT-Live-1, a multimodal speech model capable of full-duplex interaction, allowing applications to process incoming audio and stream responses simultaneously. The underlying model powers voice features in ChatGPT and is available to developers at $0.05 per minute. Developers can pair the audio interface with different reasoning backends to balance operational costs, latency, and task complexity.
According to OpenAI benchmark data, GPT-Live-1 scores 80.1 percent in full-duplex interactivity tests, up from 45.4 percent achieved by GPT-Realtime-2.1. Turn-taking latency was reduced from 1.4 seconds to 0.8 seconds, while tool-calling accuracy reached 87 percent compared to 60 percent in previous generations. In domain-specific evaluations, the model registered a 32 percent pass rate on banking voice support tasks.
The API release includes twelve new voices covering multiple accents and languages, alongside built-in automatic speech recognition (ASR) transcripts and text outputs. Early enterprise adopters, including Yelp, are integrating the API into automated phone reservation systems to manage customer interactions.
Why it matters
Full-duplex capability and 0.8s latency enable significantly more realistic, conversational voice agents for customer-facing enterprise applications.
At $0.05 per minute, voice integration costs remain high, requiring developers to carefully optimize backend model routing.
Higher tool-calling accuracy (87%) makes voice models far more reliable for executing structured backend workflows during live calls.
Source: the-decoder.com
