OpenAI releases gpt-realtime-2.1: p95 voice latency cut by at least 25%

Anyone who has tried an AI voice assistant knows the telltale flaw: the unnatural pause between the end of your question and the start of the answer. In a real phone call, half a second of extra silence is enough to make the other side sound like a robot. It is the main reason many business voice bots feel frustrating.

On July 6 OpenAI updated its voice models with gpt-realtime-2.1 and the cheaper 2.1-mini variant. The headline is not a bigger model but an optimization: in the worst cases (the notorious 95th percentile, i.e. the slowest responses) the wait drops by at least a quarter. In practice, fewer awkward moments where the user wonders if the call dropped.

There are also very practical improvements for real-world use: the model is better at understanding alphanumeric codes spoken aloud (order numbers, phone numbers, customer IDs), handles silence and background noise more gracefully, and stops more naturally when the user starts talking over the answer.

For everyday users, the effect will show up in phone-based customer service, automated switchboards, and in-app assistants built on this technology: smoother conversations and fewer rounds of "could you repeat that code, please?".