gpt-realtime-2

`gpt-realtime-2` is OpenAI's newest realtime speech-to-speech foundation model — same unified audio-in / audio-out product as gpt-realtime, with significantly better text reasoning for tool-driven voice agents. 128K text context. Voice-agent only (WebSocket).

Audio in/out rates match gpt-realtime ($32/$64 per 1M); text output is $24/1M (vs $16 on the 1.0). On CallMissed it bills against active call minutes — roughly $0.375/min — so the per-minute cost on a typical 50/50 conversation is identical to gpt-realtime.

Use it when you need top-of-the-line realtime quality plus heavy tool/function calling.