OpenAI adds realtime voice models

OpenAI is pushing voice past “transcribe then respond” into a more agentic layer where the model can reason, recover, and act while people keep talking. For builders, that shifts the bottleneck from raw latency to real-world reliability across noisy, multilingual, interruption-heavy conversations.

• –GPT-Realtime-2 adds GPT-5-class reasoning, 128K context, and better tone control for live voice agents
• –GPT-Realtime-Whisper gives developers streaming speech-to-text for captions, notes, and continuous understanding
• –The India angle matters: better transcription and translation should help multilingual voice products handle regional accents and phonetics more cleanly
• –OpenAI is also signaling production readiness with tool transparency, recovery behavior, and tighter safety guardrails
• –This strengthens the case for voice as a primary app interface, not just a front-end to text chat