Voice left the booth for a billable API
OpenAI's Realtime stack went generally available, so a voice agent can sit on a number you can forecast. GPT-Realtime-2, a translate model, and Whisper in the same family. Mistral Medium 3.5 is the open-weight brain beside it. The month is about talking systems, not prettier clips.
31 May 2026Team COEY
OpenAI's Realtime models are the thing this month actually shipped. Not a keynote voice that lives in one consumer app. GPT-Realtime-2, a dedicated translate model, and a Whisper cut, on the API, generally available. You can put a number on a conversation. You can interrupt it. You can hand it a tool. That is a product a marketer can buy without pretending they just hired a receptionist.
We already wrote the operational case in GPT-Realtime-2 Makes Voice Agents More Operational. The month-end version is the one you tell finance. The demo is over. The meter is on. If the call still sounds like a hold message, that is your brief, not the model's childhood.
A voice that cannot be interrupted is a voicemail. A voice that can take a tool call is a coworker with a headset.
The conversation you can meter
Realtime is a streaming loop. Audio in. Audio out. Text if you want a transcript for the CRM. Function calls if you want it to check a slot or pull a SKU. The old pattern was "transcribe, think, speak," three products taped together with a pause that made people hang up. This is one pipe.
GPT-Realtime-2 is the talker. It holds a thread. It does not need you to restart the personality every turn. Translate is the booth. Same pipe, another language, without booking a human for the first pass. Whisper is still the ears when you only need text.
None of this is a weights drop. You do not own the voice. You rent the conversation. That is fine if the work is a line, a booking, a status check. It is not fine if the work is a brand read you have to keep for five years.
A rented conversation can still be a good conversation. The honesty is about the archive. When the vendor changes the voice, your old spots do not get a vote. If that matters, you need a designed voice you hold, or a person.
Can you automate it
Yes, if the call returns a schema. Intent, slot, a transcript, a hang-up reason. An agent can log that. A person still has to hear the ugly ones. Do not let a model apologize its way through a refund.
Write the three tools it is allowed to call before you write the personality. Inventory. Booking. Handoff to a human. If the list is longer than that on day one, you are designing a novel, not a line.
| Piece | You keep it | You call it |
|---|---|---|
| GPT-Realtime-2 | Never | OpenAI API |
| Translate / Whisper | Never | OpenAI API |
| Medium 3.5 | Weights, or the API | Mistral |
The open brain beside the headset
Mistral Medium 3.5 is the other headline on this desk, and it is not a closed cousin. A 128B open-weight model. A 256K window. Agents in Le Chat and Vibe if you want the product. Weights if you want the box. We covered it when it landed: Mistral Medium 3.5 Wants to Be the Open Agent Model Adults Can Actually Deploy. It is not Mistral 3 from last December. Do not mix the names. Medium is the daily driver. Good enough for briefs, routing, and the first draft of a script.
Use it next to Realtime, not instead of it. One thinks. One talks. If you ask the voice model to write the brand book, you will get a brand book that sounds like a phone tree. If you ask the text model to smile into a headset, you will get a pause and a hang-up.
That pairing is the month. A mouth you can invoice. A brain you can download. Most stacks pick one and then wonder why the other half feels dumb.
What a marketer can actually plug in
A number. A script. Three tools the model is allowed to call. A human on the ugly path. That is a voice workflow. A chatbot with a speaker icon is not.
The script is still a brief. Tone, offers, the things it must never say. If you cannot write those on a page, the model will improvise, and improvisation is how brands get quoted in screenshots they did not approve.
What did not change
Video was quieter, which is a gift. LTX-2.3 and Gemma 4 from the last two months are still the local picture and planner. This month did not replace them. It gave the front door a mouth.
Hosted video vendors will tell you the clip can talk now. Sometimes it can. That is not a call center. Do not confuse a talking hero shot with a system that can book a table.
ElevenLabs and the rest of the specialist voice houses are still there. Use them when you need a designed voice, not a conversation. Realtime is the loop. A designed voice is a performance. Both are useful. They are not the same product.
Voxtral from last month is still the small local scratch voice, with the licence caveats that came with it. Realtime did not retire it. Realtime retired the excuse that you could not meter a conversation.
The headset is still a vendor
The conversation, if you pick OpenAI. The specialist voices. The talking clips. Medium 3.5 is the exception if you serve it. If you only call it through Le Chat, you are renting the product and keeping an option.
Options are good. Options are not a deployment. A weight you have not served is a bookmark with a licence.
Keep June on the local half
One Realtime path for anything that has to talk back. One Medium-class brain for the text around it, on your metal if the work is sensitive. Keep Gemma or Qwen for the work that cannot leave the building. Do not rebuild the video stack because a voice demo was charming.
A headset is not a personality. A transcript is not a strategy. A person still writes the offer.
GroundSlate is still the local half. The clip, the still, the brief. The phone tree is a vendor. This month made that vendor honest. Treat it like a headset, not like a personality.