The era of stringing together three separate models just to build a voice assistant is officially over. OpenAI moved its Realtime API out of beta today, giving every paid developer access to the same speech-to-speech tech that powers ChatGPT’s Advanced Voice Mode.
For years, the playbook was clunky but familiar: grab audio, run it through Whisper for transcription, feed the text to a language model, then pipe the response into a text-to-speech engine. The result was a voice bot that sounded like it was always catching its breath — accent stripped out, emotion flattened, latency piling up at each handoff. The Realtime API collapses all that into a single WebSocket connection. Developers stream audio in and get audio back, with the model handling interruptions natively instead of treating them as errors.
What makes this more than just a latency fix is function calling. A voice agent built on this API can actually do things — pull up a customer’s order history mid-sentence, place a delivery order, or toggle settings in an app based on a spoken command. Healthify, a nutrition and fitness coaching app, and Speak, a language learning platform, have been testing the API early. Healthify’s use case is particularly interesting: their voice coach, Ria, now understands conversational context well enough to personalize nutrition advice on the fly without making users repeat themselves.
Pricing is worth scrutinizing. Text tokens run $5 per million for input and $20 per million for output — standard GPT-4o territory. But audio tokens are a different beast entirely: $100 per million for input and $200 per million for output. OpenAI estimates that works out to roughly six cents per minute of audio coming in and twenty-four cents per minute going out. For a customer service bot fielding hundreds of calls, that math adds up fast. OpenAI also confirmed the same safety infrastructure that guards Advanced Voice Mode — automated monitoring, human review of flagged content, and the restrictions detailed in the GPT-4o System Card — applies here. Developers are required to disclose that users are talking to an AI, and using the API to spam or deceive is explicitly prohibited. Whether the cost curve drops as adoption scales is the question nobody at OpenAI is answering yet.
