Realtime voice agent pricing, estimated per minute

A realtime voice agent talks and listens at the same time over one open connection, and there are two ways to build it. A speech-to-speech model takes audio in and produces audio out directly, with no text step you control in between. A cascade chains three separate services: speech-to-text transcribes the caller, a language model writes the reply, and text-to-speech reads it aloud. Realtime APIs from OpenAI, Google and Amazon are the first kind. This page prices those, plus the bundled platforms that sell a whole agent for a flat rate.

The billing units differ by vendor. Speech-to-speech models are billed in audio tokens, quoted per million: input tokens for what the caller says and output tokens for what the model says, with a lower rate for cached input on some models. Platforms such as xAI, Deepgram, ElevenLabs and Ultravox quote one price per minute of conversation instead. A per-million-token price cannot be compared with a per-minute price directly, so every row in the table is turned into an estimated dollars-per-minute figure.

The table covers 10 offerings: 5 billed in audio tokens and 5 billed at a flat per-minute rate. Among the token-billed models, estimates run from $0.0081 per minute for Amazon Nova 2 Sonic to $0.048 for OpenAI gpt-realtime-2.1. The cheapest flat rate is $0.05 per minute from OpenAI gpt-live-1.

Vendors do not publish how many tokens a minute of conversation uses, so the estimate rests on a stated assumption. Audio is tokenised at a roughly fixed rate: about 10 input tokens per second of caller audio and about 20 output tokens per second of model audio. Over a full minute that would be 600 and 1,200 tokens. We assume a 50/50 split of talk time, so each side is active for half the minute. That gives 300 input tokens and 600 output tokens per session minute.

The formula is (300 × input price + 600 × output price) / 1,000,000, with both prices in dollars per million audio tokens. For a model priced at $32 per million input and $64 per million output, that is (300 × 32 + 600 × 64) / 1,000,000, or $0.048 per minute. The constants are stored with the data: 300 input and 600 output audio tokens per session minute.

Some vendors publish their own per-minute figure for a token-billed model, and where one exists the table uses it instead of our arithmetic. Those rows are marked with a tilde and a tooltip. Flat-rate rows use the published price as it is. Every row carries a confidence tag showing whether the price was read from the provider, taken from the LiteLLM community price list, or reported by a third party.

Treat every token-based figure on this page as an audio-only floor, not a forecast of your bill. The formula counts the audio that crosses the wire once. A production agent pays for more than that, and real bills are often two to four times the estimate.

The biggest reason is context re-billing. A realtime model reprocesses the conversation so far on each turn, so a five-minute call costs more than five times a one-minute call, because earlier turns are fed in again as input. Cached input is cheaper where a vendor offers it, but not free. System prompts, tool definitions and retrieved documents are billed as text tokens on top of the audio. Every tool call adds a request and a response, and interruptions can leave audio that was generated but never heard billed anyway. Silence and hold time count toward session minutes on most platforms.

Flat per-minute vendors fail in a different way: the headline price often covers the voice pipeline only. 2 of the 10 rows in the table are marked as not including the language model, so the model behind the agent is billed separately at its own token rates. Even where it is included, some platforms charge extra for telephony, concurrency or recordings. For a closer budget, estimate the model and tool-call side with the agent cost estimator, which handles repeated context across many steps.

At 10,000 minutes a month, the lowest estimate in the table (Amazon Nova 2 Sonic at $0.0081 per minute) comes to $81.00, and the highest (xAI Grok Voice 3.0 at $0.08) comes to $800. Multiply by two to four before committing to a budget, and measure a real session before you scale.

The cascade card above prices the do-it-yourself alternative as a floor: the cheapest streaming transcription rate on the speech-to-text page plus the cheapest text-to-speech rate on the text-to-speech page, for the half of each minute in which the agent speaks. It leaves out the language model entirely, so it is not directly comparable with a speech-to-speech row. It is the part of a cascade that you can look up. Add the model tokens for your prompt and conversation length to finish the comparison.

Beyond price, the trade-offs are practical. A cascade lets you swap each stage, log the transcript, inspect and filter the text before it is spoken, and use any language model. Speech-to-speech models skip the transcript step, which can lower latency and keep tone and interruptions natural, but give you less control over wording. Neither approach is cheaper in general. It depends on conversation length, prompt size and how much of the call is silence.

Only price is compared. We do not score voice quality, latency, accuracy, language coverage, interruption handling or tool-calling reliability, and a cheaper model that mishears callers costs more once you count failed calls. Volume discounts, committed-use contracts, free tiers and promotional rates are left out unless they are the published list price. Telephony charges for phone numbers and call minutes are separate from every figure here.

Prices change often in this category and several models are still labelled previews. Treat the confirmation date on each row as part of the data, and confirm the rate on the vendor's own page before you sign anything.

Prices are refreshed daily from the LiteLLM community price list for models it covers, and read by hand from vendor pricing pages for the rest. Token-billed rows are converted with the formula above, and flat rows are shown as published. Estimated rows are marked, and rows from resellers or third parties are tagged and can be hidden with the verified-only filter. The assumptions are part of the published data file, so a change to them shows up as a change to the data.

Sources used for the current rows: github.com (opens in new tab), deepgram.com (opens in new tab), elevenlabs.io (opens in new tab), ai.google.dev (opens in new tab), developers.openai.com (opens in new tab), ultravox.ai (opens in new tab), docs.x.ai (opens in new tab).

Not listed: Hume EVI (no official price) and Azure Voice Live (no official price). If you have an official price for a missing product, send it to us with the link.