How Realtime Voice API Pricing Works

Audio is a different token type, priced on its own schedule

A text chat and a voice call to the same model family don't just differ in how they feel to use — they're billed on entirely separate rate cards. Realtime APIs like OpenAI's gpt-realtime and Google's Gemini Live convert spoken audio into tokens the same way text becomes tokens, but the audio token type carries its own, much higher per-million price. The result: a voice conversation that sounds like a two-minute chat can bill more than a long text exchange, because every second of speech — yours and the model's — is metered at a rate text never touches.

The actual gpt-realtime rate card

OpenAI publishes separate per-million-token rates for text and audio on its Realtime API, and the gap between them is the whole story:

Audio input costs 8x what text input costs on the same model. Audio output — what the model speaks back — costs 4x audio input and 16x text output. Both directions of a voice call are billed, and the direction the model talks is always the pricier one, which matters because a chatty agent costs more per turn than a terse one, independent of what the caller said.

How many tokens is one minute of talking?

There's no single published conversion, because it depends on the audio codec each model uses internally, but working estimates across speech-to-speech models cluster around 1,000 to 1,600 audio tokens per minute of speech. A five-minute call where the caller speaks for roughly half of it and the agent replies for the other half turns into audio token volume on both sides before a single word of text instructions is added — and that volume scales with minutes, not characters, so a caller who talks slowly or pauses often still burns the same tokens per minute as one who speaks quickly.

Worked example: one 10-minute call, three models

Same call — 5 minutes of caller speech (audio input), 5 minutes of agent speech (audio output), at 1,150 audio tokens per minute (the working mid-estimate used by our calculator):

Scaled to 2,000 calls a month at that same 10-minute length — 11.5 million audio-input tokens and 11.5 million audio-output tokens a month — the gap compounds: $1,104/mo on gpt-realtime, $345/mo on gpt-realtime-mini, and $173/mo on Gemini's native-audio model for the identical call volume. Picking the mini tier alone is a 3.2x saving; picking the cheapest of the three is a 6.4x saving over the flagship — before touching call length or volume at all.

The one lever that cuts the bill without changing the experience

Cached audio input on gpt-realtime is priced at $0.40 per million tokens — 80x cheaper than the $32 fresh-audio rate. It only applies to audio the model has already processed earlier in the same session (a repeated opening script, or prior turns still held in context), not to new speech, so the saving is proportional to how much of a call is genuinely repeat content. On a system that reuses a fixed audio preamble across every call, routing that preamble through the cache instead of resending it as fresh audio turns a $32/million-token cost into a $0.40 one for that portion — the same caching logic that makes memorizzazione nella cache tempestiva valuable for text, applied to audio.

Errori comuni

• Budgeting voice features at text-API rates. Reusing a per-token cost estimate from a text chatbot for a voice feature undercounts the bill by 4-8x before the agent's speech is even added.
• Ignoring agent talk-time as the main cost driver. Audio output is priced above audio input on every realtime model published so far, so a verbose agent that over-explains costs more per call than a terse one answering the identical questions — trimming agent response length is a direct lever on the largest line item.
• Assuming a cascaded pipeline is always worse. Speech-to-text → text LLM → text-to-speech moves most of a conversation through cheap text tokens and commoditized transcription/synthesis pricing. It loses the low-latency, natural-turn-taking feel of a native speech-to-speech model, but for transactional, cost-sensitive use cases it's frequently the cheaper architecture per minute.
• Not separating input and output minutes when estimating cost. Because the two directions are priced differently (and asymmetrically across models), a call where the agent talks twice as long as the caller costs meaningfully more than the reverse, even at identical total call length.

Continua ad imparare

Domande frequenti

Why is realtime voice API pricing so much higher than text chat?

Because speech-to-speech models bill audio as its own token type, priced separately from and well above text tokens. On OpenAI's gpt-realtime, audio input is $32 per million tokens versus $4 for text input (8x), and audio output is $64 per million versus $16 for text output (4x). A voice call that sounds short can carry far more billable tokens per minute than the same exchange typed out, because speech is a denser, higher-rate signal than text.

How many audio tokens does one minute of speech use?

Roughly 1,000 to 1,600 audio tokens per minute, depending on the model and its audio codec. Both directions are billed independently: the minutes the caller speaks are audio input tokens, and the minutes the agent speaks back are audio output tokens, usually at the highest rate on the bill. A ten-minute call with both sides talking can therefore carry on the order of 10,000-16,000 billable audio tokens before any text instructions are added.

Does cached audio input actually save money?

Yes, dramatically, where it applies. OpenAI prices cached audio input at $0.40 per million tokens versus $32 for fresh audio input on gpt-realtime — an 80x difference. Caching only helps for audio segments the model has processed before in the same session (such as a repeated system clip or earlier turns still in context), not for new speech, so the saving depends entirely on how much of a session's audio is actually repeat content.

Is a cascaded STT + LLM + TTS pipeline cheaper than a native realtime model?

Usually, per minute. A cascaded pipeline — speech-to-text, then a text model, then text-to-speech — moves most of the exchange through cheap text tokens and commoditized transcription/synthesis pricing, while a native speech-to-speech model pays the audio token rate on both directions for the full conversation. The tradeoff is latency and naturalness: native realtime models keep tone, interruption handling and timing that a three-step cascade reintroduces delay into. Use realtime where conversational quality justifies the premium, and cascade where the interaction is more transactional or cost-sensitive.