GPT-4o was the only OpenAI model trained natively on voice — predicting the next sound, not converting speech to text. The new GPT-5.5 Real-Time brings that voice-native training to the GPT-5 class. Ben says this is the most exciting voice AI development in years.