OpenAI Realtime API

Our full take

The Realtime API is not a voice-agent product you log into and configure. It is the engine underneath one. OpenAI gives you a single speech-to-speech model, meaning one model that hears the caller, works out what to say, and speaks back, without bolting together a separate speech-to-text step, a language model and a voice. That tight loop is why it sounds so good. The current flagship is gpt-realtime-2, and its voices, especially the two newer ones called marin and cedar, are about as natural as anything you can buy today.

So that is the upside, and it is a real one. Speech quality and the OpenAI ecosystem. If you want an agent that handles interruptions gracefully and does not have that tell-tale robotic seam where transcription hands off to the model, this is the strongest starting point on the market. It also speaks 32-plus languages with native prosody (the rhythm and stress that make speech sound local rather than translated), and it does tool calling and MCP, so it can actually look things up and take actions mid-call.

Now the catches, and there are three worth knowing before you compare it to a Vapi or a Retell.

First, you build the agent yourself. There is no dashboard, no flow builder, no campaign manager. You get an API and SDKs (Python and JavaScript), and you write the code that connects a phone call to the model. For a developer that is a day or two of work. For a non-technical buyer it is a project you need to hire for.

Second, there is no phone line. The Realtime API connects over WebRTC, WebSocket or SIP, but it does not give you a number people can ring. You bring your own telephony, almost always Twilio, and pay Twilio separately, usually around $0.014 a minute on top. That is the same trade Deepgram and Cartesia ask of you, and it is fine once you know it is coming.

Third, and this is the one that trips people up, the billing is by audio tokens, not by the minute. This is genuinely harder to predict than a flat per-minute rate, so let me show the workings.

OpenAI’s pricing page lists gpt-realtime-2 at $32.00 per 1M audio input tokens ($0.40 per 1M if the input is cached) and $64.00 per 1M audio output tokens. Text context, your system prompt and the running conversation, is billed separately at $4.00 per 1M input ($0.40 cached) and $24.00 per 1M output. There is also a cheaper gpt-realtime-mini at $10.00 in, $0.30 cached, $20.00 out per 1M audio tokens, for when top quality is not the point.

To turn tokens into minutes you need OpenAI’s encoding rule: input (caller) audio is one token per 100 milliseconds, so 600 tokens a spoken minute, and output (agent) audio is one token per 50 milliseconds, so 1,200 tokens a spoken minute. Take a minute of conversation split roughly half caller, half agent. That is about 300 input audio tokens (about $0.01) plus 600 output audio tokens (about $0.04), so the raw audio is only around $0.05 a minute.

Here is the honest part. That $0.05 is the floor, not the bill. The expensive bit is the text context, because your system prompt and the whole conversation so far get re-sent on every turn, and a chatty agent with a long prompt racks those text tokens up fast. An independent breakdown across 11 call profiles landed at roughly $0.18 to $0.46 a minute with no caching, dropping to about $0.05 to $0.10 a minute once prompt caching is switched on. Caching matters a lot here, so if you build on this, build it in from day one. The headline on this page is an estimate of $0.18 a minute, not a published rate, and the all-in band is wide on purpose because your prompt design genuinely moves the number.

On compliance the Realtime API is on solid ground, because it inherits the OpenAI API platform’s posture. OpenAI’s enterprise-privacy and trust pages state SOC 2 Type 2 for the API, the ability to sign a HIPAA Business Associate Agreement, GDPR support with a Data Processing Addendum, and ISO 27001 and 27701. That is enough paperwork for a regulated build, though as always you get the BAA signed before you put protected health data anywhere near it.

My read. Pick the Realtime API if voice quality is the thing you will not compromise on, you are happy to write code, and you can stomach a per-minute cost that you have to model rather than read off a page. Pick a packaged platform instead if you want the phone line, the dashboard and a predictable bill handed to you. The voice is the best reason to choose this. The token pricing and the do-it-yourself assembly are the reasons plenty of teams do not.

The 1 to 10 scores on this page are an editorial preview, our provisional read to get the framework in place, not a measured result. We have not run the Realtime API through our own call tests yet, so there is no Voxrater latency figure here. The pricing, voice and compliance detail is sourced from OpenAI’s pricing, realtime and enterprise-privacy pages plus one independent cost analysis, captured 2026-05-31.