OpenAI Releases Three Realtime Audio Models

OpenAI released three new realtime audio models on Thursday, advancing voice-based AI toward more capable, conversational agents that can reason, translate, and transcribe speech on the fly. The models, GPTRealtime2, GPTRealtimeTranslate, and GPTRealtimeWhisper, are now available through the company's Realtime API and can be tested in its developer playground.

At the core is GPTRealtime2, OpenAI's most advanced voice reasoning model, designed to handle complex spoken interactions. It maintains context across longer sessions with a 128,000-token window, manages interruptions, calls external tools in parallel, and adjusts tone for different scenarios, such as calm problem-solving or empathetic responses. The model uses short preambles like "one moment" to buy thinking time and provides audible feedback on actions, such as "let me check that." Developers can set reasoning intensity from minimal to extra high, with benchmarks showing improvements over predecessors, including 96.6 percent accuracy on Big Bench Audio at the high setting.

OpenAI described the model as bringing reasoning power comparable to its latest large language systems, enabling patterns like "Voice-to-Action,"...