What is gpt-realtime, and what changes now that the Realtime API is out of beta?

Yesterday, August 28, 2025, OpenAI announced two things for developers who build voice applications. Its Realtime API, the service that lets apps hold spoken conversations with an AI model, is now generally available, meaning it has left beta and is supported for production use. And it released a new model behind it, gpt-realtime, which OpenAI calls its most advanced speech-to-speech model so far.

You will not find a new button in ChatGPT because of this. The news is for the companies building phone assistants, support lines, tutoring apps and other voice agents, AI systems that listen, talk and take actions. If you call a business in the coming months and a natural-sounding assistant answers, it may well run on tools like these.

What OpenAI released

The Realtime API first appeared in public beta last October. OpenAI says thousands of developers have used it since and helped shape this release.

How speech-to-speech works

Many voice assistants are built as a pipeline of three separate models: one turns your speech into text, a second writes a text reply, and a third reads that reply aloud. Each hand-off adds delay, and the middle model never hears how you said something, only the words.

A speech-to-speech model does the whole job in one step. It listens to the audio and produces audio in return. That cuts the waiting time and keeps details that text loses: a laugh, hesitation, a change of tone. OpenAI says gpt-realtime can pick up such non-verbal cues, switch languages in the middle of a sentence, and follow instructions about style, such as speaking quickly and professionally or warmly and patiently.

Two improvements target real business calls. The model is better at reading required scripts word for word, such as a legal disclaimer, and at catching and repeating strings of letters and numbers, like phone numbers or vehicle identification numbers, including in Spanish, Chinese, Japanese and French.

OpenAI reports these results against its previous voice model from December 2024:

Tool calls also no longer stall the conversation. If a lookup takes a while, the model can keep talking while it waits for the result.

The new connections, in plain terms

SIP (Session Initiation Protocol) is the standard that internet phone systems use to set up calls. Support for it means a voice agent can answer or place an ordinary phone call without extra telephone software in between.

MCP (Model Context Protocol) is an open standard for connecting AI models to tools and data. With a remote MCP server, a developer adds a server's web address to the session, and the API handles the tool calls automatically. Pointing the session at a different server adds different abilities.

Image input lets a user share a picture during the conversation and ask "What do you see?" or "Read the text in this screenshot." The app decides which images to share and when.

Here is a worked example of the instructions a developer might give a voice agent for a dental clinic:

You are the phone assistant for Maple Street Dental. Speak warmly and at a calm pace.
At the start of every call, read this sentence exactly: "This call is answered by an
AI assistant and may be recorded."
When a caller gives a phone number or booking code, repeat it back digit by digit
and ask them to confirm.
If the caller switches to French or Spanish, continue in that language.
Use the check_availability tool before offering any appointment time.
Never give medical advice; offer to book a visit instead.

Each line maps to one of the improvements above: tone, word-for-word scripts, alphanumerics, language switching and tool use.

Safety, privacy and price

OpenAI runs classifiers on Realtime sessions and can halt conversations that break its harmful content rules. Its policies require developers to tell people they are talking to an AI unless it is obvious, and the API uses only preset voices to make impersonating real people harder. EU-based apps can keep their data in Europe.

Audio going into the model now costs $32 per million tokens, or $0.40 when the input is cached (already processed and stored from earlier in the session), and audio coming out costs $64 per million tokens. A token is a small unit of data the model processes; for audio, it represents a short slice of sound. Developers also get finer control over how much of a long conversation the model keeps, which can lower the cost of long calls.

How can a developer build a first voice agent with gpt-realtime?

Start in OpenAI's Playground, where you can talk to gpt-realtime, try the Marin and Cedar voices and adjust instructions before writing code. OpenAI also publishes Realtime API documentation and a prompting guide specific to voice. Begin with a narrow job, such as answering opening hours or taking a message, and write instructions like the example above.

Next, add one tool, such as a calendar lookup, and test how the agent behaves when the tool is slow or returns nothing. Only then connect a phone number through SIP or an MCP server for more tools. Record test calls with colleagues, listen for mistakes in names and numbers, and keep a clear way for callers to reach a human. Voice agents talk to the public, so disclosure, confirmation steps and careful testing matter as much as a natural voice.