OpenAI has launched three brand-new voice models via its Realtime API, with the core product GPT-Realtime-2 integrating what OpenAI calls 'GPT-5 level reasoning capabilities' for the first time. Unlike previous models that only offered text-to-speech and basic conversational functions, the new models are designed to handle multi-step tasks in real time, call external tools in parallel during conversations, and act as collaborative agents that manage interruptions. Also released were the streaming translation model GPT-Realtime-Translate and the real-time transcription model GPT-Realtime-Whisper. This release has adjusted the financial and technical parameters, reshaping the cost structure of the API market. GPT-Realtime-2 offers five levels of reasoning intensity adjustment. Under the 'high' intensity setting, its accuracy in the Big Bench Audio test increased from 81.4% in the previous generation GPT-Realtime-1.5 to 96.6%. In terms of cost, GPT-Realtime-2 audio input is priced at $32 per million tokens, but cached input fees have dropped to $0.40. The auxiliary model GPT-Realtime-Translate enables real-time conversion from over 70 input languages to 13 output languages, priced at $0.034 per minute. Whisper transcription services are priced at $0.017 per minute.