What Realtime Voice APIs Cost per Minute of Conversation

The sticker rate is not the bill

Two different pricing models are in the market and they look nothing alike on an invoice. OpenAI, Google, Microsoft and Amazon price realtime voice by the audio token. ElevenLabs, AssemblyAI, Retell, Cartesia and Deepgram wrap one or more of those models and charge a single per-minute rate on top. Both can be the right answer. Neither is cheap by default.

The reason token pricing bites is that audio tokens are small and conversations are long. OpenAI's Realtime models tokenize user audio at one token per 100 ms and assistant audio at one token per 50 ms, which works out to 600 input tokens and 1,200 output tokens for a minute of each side speaking (Forasoft analysis). Those numbers are not large per minute. They become large because every subsequent turn replays the whole prior conversation as new input unless caching is on.

Per-minute bundles look simpler but hide the LLM, the telephony and sometimes the TTS. ElevenLabs charges $0.08 per minute for every plan on their Conversational AI, with the LLM billed on top. AssemblyAI bundles STT, LLM, TTS and hosting into $4.50 an hour, or $0.075 per minute of open session time. Which one is cheaper depends entirely on how chatty your agent is and whether you talk to a hosted model or bring your own.

What a minute of OpenAI Realtime actually costs

OpenAI's flagship gpt-realtime-2.1 lists at $32 per million audio input tokens, $64 per million audio output tokens, and $0.40 per million for cached audio input (OpenAI pricing). The mini variant, gpt-realtime-2.1-mini, lists at $10 input, $20 output and $0.30 cached. There is no per-minute line item in OpenAI's price list; the $0.10 figure that circulates is a rough all-in estimate.

The underlying math: user audio is one token per 100 ms, assistant audio is one token per 50 ms (Forasoft). A full minute of the user speaking into the API is 600 input tokens; a full minute of the assistant speaking back is 1,200 output tokens. At flagship rates that is $0.019 heard and $0.077 spoken. Split a minute 50/50 and the first turn runs about $0.048 in pure audio.

The replay is where it gets expensive. Each new assistant response gets the whole conversation so far as input again. Hackernoon's measurements across 4,000 sessions put gpt-realtime-mini at $0.063 to $0.146 per minute in production, with short cached calls landing near $0.02 to $0.05 and chatty uncached calls climbing to the $0.146 ceiling. Turn caching on and the input replay line drops from $32 to $0.40 per million, an 80x cut.

In short: gpt-realtime-2.1-mini lands at $0.02 to $0.15 per minute in production. The flagship runs on token rates roughly 3x higher, so plan on $0.10 to $0.30 per minute before caching discipline and closer to $0.10 after, consistent with the $0.10 per-minute figure on the RankGrids ranking.

Which speech-to-speech API is cheapest right now

At published 2026 rates, the cheapest native speech-to-speech models are Google's Gemini 3.8 Live and Amazon's Nova 2.5 Sonic, both well under three cents a minute before replay. Google quotes Live API audio input at $3.00 per million tokens or $0.005 per minute, and audio output at $12.00 per million or $0.018 per minute. That is roughly $0.023 for one minute of a half-and-half conversation.

Amazon's Nova Sonic lists at $3 per million audio input tokens and $12 per million audio output tokens through Amazon Bedrock, which works out to around $0.015 per metered minute on an even split. It is the cheapest sticker rate on the list, but the 1,080 ms end-to-end latency in the RankGrids ranking pushes it toward asynchronous use cases rather than live phone work.

Among managed bundles, the cheapest at face value is Cartesia's voice agents at $0.03 per minute on all tiers, with telephony at an extra $0.014 per minute if you use Cartesia-provided numbers. That is under half of ElevenLabs Agents at $0.08 and under half of Deepgram's standard voice agent tier at $0.075.

One wrinkle worth flagging. The cheapest API for a prototype is almost never the cheapest API in production. Free tiers burn off in a day of traffic. Cached rates make big models surprisingly competitive once you get past the first thousand conversations. And LLM costs, which a lot of managed bundles do not include, swing the total by 2x to 3x depending on model choice.

How to estimate a voice agent's cost before you ship

The usable formula, adapted from AssemblyAI's voice agent cost guide, has four lines. For any one call: audio in, audio out, LLM tokens, telephony. Everything else is a surcharge on one of those.

Audio in is user speech time times your input rate. Audio out is assistant speech time times your output rate. In a typical support call, humans talk about 40 percent of the time and the agent talks the other 60 percent, though research tools and real-time banking skew closer to 30/70 agent-heavy. Multiply both by your chosen model's token rates.

The LLM line is the one that trips people up. On a pure speech-to-speech model like gpt-realtime or Gemini Live, there is no separate LLM bill; the audio tokens are the tokens. On a cascade like ElevenLabs Agents, Retell, Deepgram Voice Agent or Azure Voice Live Standard, you still need to add the input-plus-output tokens for whichever text LLM you plugged in. Retell's own calculator opens at $0.11 per minute broken out as $0.055 voice infra, $0.04 LLM and $0.015 TTS to show how the layers stack.

Telephony is per-minute and roughly $0.01 to $0.015 if you bring Twilio or similar, free on Retell's bundled SIP, and $0.014 on Cartesia-provided numbers. If you need real SIP trunking with a 99.9 percent SLA, that line stops being optional.

The pragmatic shortcut: multiply your target per-minute rate by expected monthly minutes, then add 30 percent for context replay on uncached sessions, 15 percent for LLM on any cascade, and roughly $0.01 for telephony. A 10,000-minute month on ElevenLabs Agents with GPT-4o-mini and Twilio therefore budgets to about $1,050, not $800.

The five places the bill surprises teams

One: context replay. Every turn resends the whole audio history as input unless prompt caching is on. On gpt-realtime-2.1 that is $32 per million audio tokens becoming $0.40 per million cached, an 80x cut on a line that would otherwise dominate past minute three of a call. If your platform does not expose cache control, assume replay will double the raw audio bill on calls over five minutes.

Two: long system prompts. A 3,000-token text system prompt resent on every turn adds about $0.012 per turn uncached on the flagship (OpenAI pricing lists text input at $4 per million tokens). Ten turns and you are at $0.12 just for the preamble, before anyone spoke. Caching the prompt drops that line to about $0.0012 per turn.

Three: silence. Open-session bundles charge for wall-clock time, not speech time. AssemblyAI's $4.50 an hour is metered on session open, which means a 30-second awkward pause is billed the same as 30 seconds of conversation. Token-metered models like OpenAI Realtime do not bill silence on the input side, because silence is not transcribed, but they will bill any assistant audio response including filler phrases.

Four: burst concurrency. ElevenLabs charges a $0.16-per-minute burst rate if you exceed your plan's concurrent-call cap, double the headline $0.08. Plans that look generous on minutes but stingy on concurrency can quietly cost double during peak hours.

Five: feature gates. Microsoft's Voice Live API runs three tiers, with native audio on the Pro tier at $44 per million input and $88 per million output. The Basic tier using GPT-4o Mini Realtime is $11 and $22 per million, exactly a quarter of Pro. Picking the wrong tier for your workload wastes most of your budget, and the pricing page does not force you to think about which one you need.

When a flat-rate bundle beats a raw token meter

A flat per-minute bundle is a bet that your conversations will be long, chatty and operationally expensive to orchestrate. A token meter is a bet that your conversations will be short, bursty or that you can out-engineer the replay problem with caching.

The crossover is roughly here: if your average session is under three minutes and you are willing to maintain caching discipline, raw token pricing on gpt-realtime-2.1-mini or Nova 2.5 Sonic beats every bundle. If your average session is eight to fifteen minutes with a stable prompt, a bundled rate from AssemblyAI or Deepgram Voice Agent is usually cheaper and much easier to forecast.

Hume EVI 3's pricing structure is a plan-based per-minute rate, cited at $0.07 on Starter, $0.06 on Pro and $0.05 on Scale, with sales-negotiated volume rates Hume describes as going well below $0.02 per minute at the highest volumes. That makes EVI 3 behave like a bundle at low volume and more like a token meter at high volume, which is unusual but useful if you have both prototype and production traffic on the same account.

The three-way comparison OpenAI gpt-realtime-2.1 vs Hume EVI 3 vs ElevenLabs Agents is a good reference point for teams choosing between these three on cost, latency and voice quality in one pass.

LLM, telephony and multilingual are the lines calculators miss

If your voice stack is a cascade, the LLM is a separate contract. Retell shows this transparently; ElevenLabs makes you pick a provider and pay them directly. On a 10-minute call with GPT-4o-mini at 500 input and 500 output tokens per minute averaged, that is roughly $0.03 of LLM per call, or 7 to 10 percent of the total. Swap to GPT-4o and that line is $0.30 per call and now matters more than the voice.

Telephony is the second forgotten line. OpenAI's Realtime API supports native SIP trunking with a 99.9 percent SLA but you still pay per-minute telephony on top. Retell includes SIP in the platform fee, which is unusual and worth money if you run outbound campaigns at scale. Gemini 3.8 Live and Hume EVI 3 ship without SIP, which means you need WebRTC or a telephony partner and another line on the bill.

Multilingual is sometimes priced as a premium and sometimes just works. Gemini Live covers 97 languages at the same rate. Alibaba's Qwen-Audio-3.1-Realtime lists at $0.018 per minute across 31 languages including 20 Chinese regional varieties, the cheapest rate on this list for serious Chinese coverage. Azure Cognitive Services hits 140 languages but at a much higher effective rate once you factor neural TTS and LLM separately.

What this all means for your build

If you are prototyping, use Gemini Live or Hume's free tier and ignore the sticker rate entirely until you have 100 real conversations to analyze. If you are shipping a consumer app with short bursty sessions, gpt-realtime-2.1-mini with caching is the cheapest credible default, landing near two to five cents a minute on the Hackernoon measurements. If you are running a contact center with long structured calls, a bundled rate from AssemblyAI or Deepgram Voice Agent is easier to forecast and usually within 20 percent of what you would pay piecing it together. If you need SIP with an SLA and native audio quality, Azure Voice Live Pro or OpenAI's Realtime API with SIP are the two credible choices and both land above ten cents a minute all-in.

The pricing-page number is almost never the number on the invoice. Use the formula, measure your actual input/output ratio, turn caching on, and budget 15 to 30 percent of slack for the lines you cannot predict in advance.