OpenAI ships GPT-Realtime-2, voice reasoning

This is OpenAI trying to make voice agents feel like native model products, not stitched-together STT plus TTS pipelines. The interesting part is less the audio novelty and more the jump in reasoning, tool use, and session memory that makes spoken workflows usable.

• –GPT-Realtime-2 adds GPT-5-class reasoning, stronger instruction following, and more reliable tool use for complex voice-agent flows
• –The model ships with a 128,000-token context window, which matters for long, interruption-heavy conversations
• –OpenAI is pairing it with realtime translation and transcription models, so developers can build full voice stacks without mixing vendors
• –Pricing is aimed at production workloads, not toys: $32 per 1M audio input tokens and $64 per 1M audio output tokens
• –The real test will be latency and turn-taking under messy, real-world speech, not benchmark hype