Voice agents have been expensive to run and painful to orchestrate, not because the models can’t handle the conversation, but because context limits forced companies to build layers of session reset, state compression, and reconstruction into every deployment. OpenAI’s three new voice models are designed to reduce that overhead and change the way engineers can think about integrating 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 reasoning, translation, and transcription into specialized components rather than bundling them into a single speech product.
The company said in a blog post Realtime-2 is its first “GPT-5 class reasoning” voice model and can handle difficult requests and keep conversations flowing naturally. Realtime-Translate understands over 70 languages and translates them into 13 others at the speaker’s pace, and Realtime-Whisper is their new speech-to-text transcription model.
These three actions are no longer contained within a single stack or model. GPT-Realtime-2 could technically handle transcription, but OpenAI is directing different tasks to specialized models: Realtime-Translate for multilingual speech and Realtime-Whisper for transcription. Enterprises can map each task to the appropriate model rather than routing everything through a single, all-encompassing voice system.
New OpenAI models compete Mistral Voxtral Modelswhich also separates transcription and focuses on business use cases.
What should companies do?
More businesses are seeing the value of voice agents now that more people feel comfortable conversing with an AI agent, and also because of the wealth of data from voice interactions with customers.
Organizations evaluating these models will need to consider their orchestration architecture, not just the quality of the model; specifically, whether your stack can route discrete voice tasks to specialized models and manage state across a 128,000-token context window.
