OpenAI's new voice stack is an agent play, not a party trick

On 7 May, three models: GPT-Realtime-2 (first voice model with GPT-5-class reasoning), Translate (live, 70+ to 13 languages) and Whisper (streaming transcription). Translate and Whisper are billed by the minute; Realtime-2 by tokens. That pricing split is the most honest thing about the launch: per-minute works for commodity streaming where cost is proportional to volume. Token pricing reflects variable reasoning depth — and OpenAI charging per token for the reasoning model is tacit acknowledgment that this is where the actual work, and the real cost, concentrates.

Translation and transcription are increasingly commodity — every major lab does them, several third-party providers do them well, and the per-minute pricing reflects the race to the bottom. The thing competitors cannot trivially copy is a voice model with frontier reasoning baked in, low-latency enough to hold a conversation. That's the component that turns voice from 'dictation with a personality' into an agent that can actually do tasks while you talk — look up your account, book the reservation, resolve the dispute — without the developer having to route audio through three separate services and pray the latency is acceptable.

Put the models side by side:

Read the defensibility column, not the count.

This dropped in early May, right before Sesame's voice app (28 May) and Apple's Siri AI reveal at WWDC (8 June). That's not coincidence. The voice-agent race is real and the developer ecosystem is up for grabs. Apple controls the microphone and the lock screen, but it can't ship an API. Sesame controls a slick consumer product but needs third-party models. OpenAI's play is to become the infrastructure underneath both — the reasoning engine that every voice-native app has to reach for because nothing else can do the request scope it enables.

Whoever owns the developer layer owns the ecosystem. The voices everyone coos over — in customer service, in AI assistants, in the apps Sesame wants to build — will run on someone's API. OpenAI moved first with the most capable stack, which means developers who start building now will build for this, and the switching costs rise with every app shipped. That's the actual product story under the press release.

The question isn't whether OpenAI shipped this. It's whether Realtime-2's reasoning quality, in production, actually closes the gap between voice demos and voice agents — or whether developers hit latency or quality walls that push them back towards chained architectures. Watch for the first wave of apps claiming to use it by Q3. If the reasoning holds, this is the moment voice agents became real. If it's shallow, it's a better streaming API with good marketing.

There's a second tell worth watching: the billing split isn't just a pricing footnote, it's a lock-in mechanism. Per-token reasoning costs scale with how much your app actually thinks, which means the more capable the agent a developer builds, the deeper the dependency on OpenAI's most expensive, least-substitutable model — exactly the inverse of the commodity translation and transcription, where you can swap to Deepgram or AssemblyAI on a price war and barely notice. So the strategic move underneath the launch is to make the cheap parts interchangeable and the valuable part sticky. Developers who architect around Realtime-2's reasoning aren't buying a feature; they're choosing a vendor for the part of the stack that's hardest to rip out later. That's the part the launch post won't say out loud — and the part a builder evaluating this should price in before committing.