OpenAI ships three next-gen voice models in its API

On 7 May, OpenAI added three audio models to its Realtime API. GPT-Realtime-2 brings GPT-5-class reasoning into a live voice model — the first time OpenAI has offered frontier reasoning in-band during a voice conversation. GPT-Realtime-Translate handles live speech translation from 70+ input languages into 13 outputs, keeping pace with the speaker. GPT-Realtime-Whisper streams transcription in real time as audio comes in. All three are accessible through a single Realtime API connection.

Translate and Whisper are billed per minute of audio — standard pricing for commodity streaming tasks. GPT-Realtime-2 is billed by token. That difference matters: token billing reflects variable reasoning depth, and it's OpenAI's signal that this is where the expensive, defensible work sits. Per-minute billing would undercharge for complex requests; token billing prices the actual compute. The pricing architecture maps to where each model is genuinely differentiated.

What's next. The real test is whether GPT-Realtime-2's reasoning quality holds in production apps. Real-time reasoning trades some depth for latency, and the launch post is not the proof. Watch for the first wave of consumer and enterprise voice apps shipping on this stack in the months ahead — that's when the capability claim becomes verifiable.