GPT-4o opens its voice to everyone

On 13 May, OpenAI announced GPT-4o, the ‘o’ standing for ‘omni’, and with it the company quietly changed the stakes of the AI race. The new flagship model can reason across text, audio and images in real time, and it will be free for everyone, not tucked behind the paid tier. In a short demonstration, two people talked to the model as if it were a colleague at a dinner table, interrupting each other, laughing, and being answered in tones that sounded almost human. The demo was polished, but the change underneath it is structural.

The technical details matter less than the feel of them. GPT-4o responds to spoken input in an average of about 320 milliseconds, which is roughly human conversational pace, and it can understand tone, detect emotion in a voice, and speak with something closer to personality than the stilted voices of previous models. It can also look at a screen through the phone’s camera and answer questions about what it sees, in real time, which turns the phone itself into an input device rather than just a text box. The model, in short, begins to behave less like a search engine and more like a presence in the room.

The pricing decision is the more interesting part. OpenAI’s most capable model, previously a reason to subscribe to ChatGPT Plus, is now given away, with the company betting that usage, habit and the inevitable enterprise demand will repay the cost. This is the classic software strategy of commoditising the front door and selling the back office, except the commodity here is the most capable language model in the world. It is a generous gesture that also happens to be a moat: once millions of people get used to a free voice companion that answers instantly, switching to a competitor’s slightly better model becomes a tax on their habits.

There are harder questions underneath the polish, and it would be dishonest to ignore them. The model’s ability to mimic a voice raises obvious risks around impersonation and consent, and OpenAI’s decision to hold back its own voice, called Sky, after public complaints, showed how quickly that line can be crossed in public perception. The real-time visual understanding, meanwhile, is a privacy surface that no one has yet written rules for: a camera pointed at a room, streaming its contents to a model, with the user’s permission given once. Multimodal models widen the attack surface for misuse even as they widen the circle of people who can use them.

There is also the question of what happens to the humans in the loop. The demo’s charm was its conversational warmth, but the same warmth, applied to thousands of automated support lines and phone trees, could make automated systems feel indistinguishable from people without being accountable like people. That is not an argument against the technology, but it is an argument for treating the smoothness with a little suspicion.

None of this should obscure what happened on 13 May. For the first time, the most advanced model a major lab produces is available, free, to anyone with a phone, in a form that talks back in real time. Whether that is the moment AI became a utility or the moment it became a companion is still an open question, but the direction is no longer in doubt. The race is no longer about who can build the smartest model; it is about who can make one that feels natural to talk to.