GPT-4o Audio Preview: What It Is and What It Can Do

TLDR

OpenAI just released a new AI model that can actually hear and speak.
Instead of just typing, you can send it audio and it talks back in a natural voice.
It uses the same smart "brain" as OpenAI's newer Realtime API (a tool for live, instant conversations).
It’s a big step up from older text-only models because it handles sound directly.
This model is a new version of OpenAI's GPT-4o that is built for spoken conversations. It is perfect for anyone who wants to build apps where you talk to an AI out loud and it talks back, rather than just typing back and forth.

Handles audio in and out: You can send it a spoken question or sound, and it generates a spoken response, rather than just turning your voice into text first.
Powered by the Realtime API: It uses the same underlying tech as OpenAI's Realtime API, which is designed for fast, fluid, live conversations without awkward pauses.

Versus older models: Previous models usually had to use a workaround—they would turn your audio into text, think about the text, and then use a separate voice generator to speak. This new model processes audio directly, making it faster and more natural.
Versus the Realtime API: It uses the exact same core model as the Realtime API, but is built to fit into the standard "chat completions" format that developers already use for text.

The announcement does not mention specific prices, context window (how much it can remember at once), or audio length limits. Developers will need to check OpenAI's official pricing page for those exact numbers.

It is a "preview": This means it is still in an early testing phase. Things might change, and it could occasionally have bugs or quirks.
Built for developers: Right now, this is released as an API (a set of tools for programmers). Everyday users won't directly use this specific model on their own; they will use it inside apps that developers create.

GPT-4o Audio Preview, Explained Simply