OpenAI has announced the general availability of its Realtime API, featuring the new gpt-realtime speech-to-speech model designed for production-ready voice agents. This release introduces critical enterprise capabilities, including remote MCP server support for tool integration, image input processing, and SIP phone calling. The gpt-realtime model significantly outperforms its predecessor, achieving 82.8% accuracy on the Big Bench Audio evaluation for reasoning and 66.5% on the ComplexFuncBench for function calling.

These improvements allow for more complex, multi-step interactions with lower latency compared to traditional chained models. The update also includes two new expressive voices, Cedar and Marin. Industry partners like Zillow are already leveraging the technology to enhance user experiences, noting the model's ability to handle nuanced, natural conversations.

This advancement streamlines the development of sophisticated voice AI for customer support and personal assistance.