OpenAI and Microsoft roll out new voice AI models and developer tools

OpenAI and Microsoft on Thursday each unveiled new voice-focused artificial intelligence models and developer tools intended to accelerate the creation of conversational agents, setting the stage for sharper competition in voice AI and broader choices for developers and enterprises. OpenAI expanded its Realtime API to general availability and released gptrealtime, its most advanced speech to speech model, while Microsoft introduced MAIVoice1, a fast speech generation model, and MAI1preview, an inhouse text foundation model now in public testing.

OpenAI said gptrealtime offers large gains in instruction following, tool calling accuracy, and speech naturalness, citing internal benchmarks that show audio understanding accuracy rising from 65.6 to 82.8. The company also added two new voices, Cedar and Marin, updated six existing voices, and enabled image inputs, Session Initiation Protocol phone calling, and support for remote MCP servers in the Realtime API. Developers can now save and reuse prompts across sessions, and OpenAI reduced Realtime pricing by about 20 percent compared with its previous realtime model, the company said. OpenAI framed the updates as aimed at making voice...