OpenAI has introduced three new audio models through its API, expanding its push into real-time voice AI for developers. The launch includes GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper, each targeting a different part of live voice interaction.

The company said the new models aim to make voice software more useful in everyday situations. That includes handling conversations while driving, navigating airports, or getting customer support without typing.

OpenAI framed the launch around a broader shift in computing interfaces. “Voice is becoming one of the most natural ways for people to use software,” the company said.

Smarter voice interactions

GPT-Realtime-2 serves as the flagship model in the release. OpenAI described it as its first voice model with GPT-5-class reasoning capabilities. The system can process harder requests, manage interruptions, and continue conversations naturally.

The model also supports live tool usage. Developers can let the AI access calendars, search systems, or other tools while speaking with users. OpenAI said the model can explain those actions in real time using phrases like “checking your calendar” or “looking that up now.”

OpenAI also expanded the model’s context window from 32K to 128K. That allows longer conversations and more complex tasks without losing context.

The company said GPT-Realtime-2 can recover more smoothly when something fails. It also better understands industry-specific terminology, including healthcare vocabulary and proper nouns.

OpenAI shared benchmark improvements tied to live voice performance. GPT-Realtime-2 (high) scored 15.2% higher on Big Bench Audio than GPT-Realtime-1.5. The xhigh version improved instruction-following scores by 13.8% on Audio MultiChallenge tests.

OpenAI’s new audio models place the company in close competition with Google’s Gemini Live. However, the latter still stands out for fast responses and stronger language support. OpenAI’s approach seems to be focusing more on making conversations feel natural during longer interactions. The new models can handle interruptions, use tools during calls, and “keep pace with the speaker,” according to the company.

Live translation features

OpenAI also launched GPT-Realtime-Translate, a real-time translation model designed for multilingual conversations.

The model translates speech from more than 70 input languages into 13 output languages while keeping pace with the speaker. OpenAI positioned the model for customer support, travel, and cross-language communication systems.

The company pointed to examples already in development. Deutsche Telekom is building voice support tools that let customers speak in their preferred language while the AI translates conversations live.

Speech-to-Text expansion

The third release, GPT-Realtime-Whisper, focuses on live transcription. The model converts speech into text while a person speaks, supporting streaming speech-to-text use cases.

OpenAI said the broader goal involves moving beyond simple voice assistants toward systems that can actively complete tasks during conversations.

For example, Zillow is developing a voice assistant that can search for homes, filter preferences, and schedule tours from spoken requests alone.

OpenAI said these models push real-time audio systems closer to agents that can “listen, reason, translate, transcribe, and take action as a conversation unfolds.”