Voice agents are expensive to run and painful to orchestrate, not because models can’t handle conversations, but because the context ceiling has forced enterprises to build session reset, state compression, and rebuild layers into every deployment. OpenAI’s three new voice models are designed to reduce that overhead, and they change how engineers can think about building voice into a larger agent stack.
GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper integrate real-time audio into the model management stack as discrete orchestration primitives – separating conversational logic, translation, and transcription into specialized components rather than bundling them into a single voice product.
The company said in a blog post that RealTime-2 is its first voice model “with GPT-5 class reasoning” and can handle complex requests and keep conversations flowing naturally. Realtime-Translate understands over 70 languages and translates them into 13 more at speaker speed, and Realtime-Whisper is its new speech-to-text transcription model.
These three actions no longer sit inside the same stack or model. GPT-Realtime-2 can technically handle transcription, but OpenAI is delegating specific tasks to specialized models: Realtime-Translate for multilingual speech and Realtime-Whisper for transcription. Enterprises can delegate each task to the appropriate model rather than routing everything through a single, omnipresent voice system.
The new OpenAI models compete against Mistral’s Voxtral models, which also separate transcription and target enterprise use cases.
what should enterprises do
More enterprises are now seeing the importance of voice agents as more people are becoming comfortable interacting with an AI agent, and also because of the richness of data from voice customer interactions.
Organizations evaluating these models will need to consider not just the quality of the models, but also their orchestration architecture – specifically, whether their stack can deliver different voice functions to particular models and manage state in a 128K-token context window.
<a href