OpenAI Launches gpt-realtime-2.1 With Lower Latency and Reasoning in Voice Conversations

OpenAI rolled out two new real-time voice models in its API on July 6, 2026: gpt-realtime-2.1 and gpt-realtime-2.1-mini. The company says response latency has dropped by at least 25 percent at the 95th percentile, which in practice means fewer noticeable pauses when talking to a voice assistant.

The new generation of realtime models replaces the previously released gpt-realtime-2. The main change concerns audio processing: the model is better at recognizing alphanumeric strings, such as order numbers or email addresses spoken aloud, and does a better job distinguishing silence from background noise and handling interruptions while a user is speaking.

Reasoning in voice conversations

The biggest addition is a configurable reasoning level for the voice models. Developers can choose from minimal, low, medium, high, or very high, depending on whether the priority is response speed or reasoning quality. The mini model defaults to the low setting, intended to keep latency low while still adding basic multi-step reasoning and tool-use capability.

This marks the first time OpenAI has brought full reasoning to the mini variant of its realtime line. Previously, the cheaper voice models were limited to simpler tasks, with no capacity for multi-step response planning. Now gpt-realtime-2.1-mini is meant to handle more complex scenarios at the same price as the previous version without reasoning.

Business use cases

OpenAI points to specific business scenarios for the new models: customer service with tool calls made mid-conversation, appointment booking with precise character-by-character data capture, in-app voice assistants handling higher call volumes, and field data collection where accurately recognizing spoken numbers and codes matters.

The full gpt-realtime-2.1 model is priced at $4 per million text input tokens and $32 per million audio input tokens, with higher rates for audio output, reaching $64 per million tokens. The mini variant is an order of magnitude cheaper, at $0.60 per million text input tokens and $10 per million audio input tokens.

Competition in the voice segment

The real-time voice model segment is gaining importance as agents that handle customers over the phone and through apps grow more popular. OpenAI has long competed in this space against other providers' voice offerings, and cutting latency is a key competitive factor here, since even small delays in a response are easily noticeable to the person on the other end and make the conversation feel artificial.

For Polish companies building voice hotlines or in-app assistants, the new models mean they can handle more complex queries without moving to the pricier full variant. That could lower the cost of deployments in call centers or booking systems, where the limiting factor until now had been the weaker reasoning quality of cheaper voice models.

Access to both models is already open to developers using the OpenAI API, with no waitlist required. The announcement was posted on OpenAI's developer community forum and in the official API changelog, without an accompanying press event.

Sources: MarkTechPost (marktechpost.com), Let's Data Science (letsdatascience.com)