OpenAI's voice models that reason while they listen

Voice AI has spent two years sounding good and reasoning poorly — fluent on the surface, brittle the moment a request got complicated. The gap wasn't the voice; it was the intelligence sitting underneath. Transcription was a separate service. Translation was another integration. Reasoning happened in a different model entirely, and stitching them together with low enough latency to feel like a conversation was the unsolved engineering problem that made real voice agents elusive. OpenAI's 7 May update is aimed squarely at that architecture problem, putting GPT-5-class reasoning inside a real-time voice model for the first time and bundling transcription and translation alongside it.

Three models, one stack. GPT-Realtime-2 is the headline: a voice model that can actually think through a harder ask mid-conversation, rather than retreating to canned replies when the request escapes a simple pattern. Unlike earlier models, where reasoning required routing audio through a separate text pipeline with noticeable lag, Realtime-2 handles it in-line. GPT-Realtime-Translate does live speech translation from 70+ input languages into 13 output languages, keeping pace with the speaker in near real time. GPT-Realtime-Whisper streams transcription as you talk, carrying OpenAI's Whisper transcription brand into the Realtime API as a managed, low-latency streaming service rather than a batch upload.

The billing split is worth noting: Translate and Whisper are billed by the minute, Realtime-2 by token usage. That pricing architecture is a quiet signal about where OpenAI sees the cost and value concentrating. Per-minute pricing makes sense for commodity streaming tasks — roughly proportional to audio volume, easy to forecast. Token pricing reflects variable reasoning depth: a complex request consumes more context, more compute, and more tokens than a simple one. OpenAI has priced each model to match its cost profile, and the token model for Realtime-2 is an implicit acknowledgment that reasoning workloads are neither flat nor cheap.

Before this update, building a voice agent that could handle a non-trivial request involved several moving parts: a streaming transcription service (Whisper or a third party), a translation layer if the user spoke another language, a reasoning model behind an API call, and a text-to-speech synthesis step before the audio came back. Each boundary added latency and a failure mode. Most production voice agents kept the request scope narrow enough to avoid the hard reasoning step entirely, which is why they still felt scripted.

Collapsing transcription, translation and reasoning into one Realtime API eliminates those boundaries. A developer today can invoke a single WebSocket connection, stream audio in, and receive audio out — with reasoning, translation and transcription handled inside the API rather than assembled in the application layer. That reduction in integration complexity is what lets a small team ship something that would previously have taken a larger one.

The comparison table below shows what the three models cover and how they're priced:

Translate and Whisper are commodity-priced; Realtime-2 is reasoning-priced.

The timing is notable. This shipped in early May, just ahead of Sesame's iOS voice-agent app — launched 28 May — and Apple's WWDC reveal of Siri AI on 8 June. The race to own the developer layer for voice is not hypothetical; it is the active battleground. Apple controls the device microphone and the user's trust. Sesame, founded by Oculus veterans, is betting on a product-layer play over the API. OpenAI, by moving early with the most capable API stack, is making a different bet: that developers building voice-native apps will reach for the most powerful toolkit, and that the ecosystem built on that toolkit becomes self-reinforcing.

The proof will be in production apps, not launch posts. Real-time reasoning always trades depth for latency — the live model won't match the deliberate reasoning of a slow text call — and the interesting question is where that latency-accuracy tradeoff sits in Realtime-2 versus a chained architecture. If it holds up in developer testing, expect a wave of voice-native products across customer support, live translation, and AI assistant categories that couldn't previously handle the request scope. If the reasoning quality is shallow in practice, the value proposition narrows to commodity streaming with a better API experience. The answer will be apparent within a product cycle.