What one real conversation costs with a voice agent (broken down)

Every few days someone in here asks what a voice agent actually costs per call, and the answer is always “like 10 cents a minute” with no breakdown. So I priced one out line by line. For context, we’re an automation agency, 8 years in, and voice agents are a big chunk of what we build now. Mostly outbound for real estate. I used a 4 minute outbound call to a US or Canadian number with about 20 back and forths. The agent talks for roughly 40% of it and runs on a system prompt of around 2,000 tokens. Pretty common stack, priced at list with no volume discounts. Vapi platform fee (4 min at $0.05) $0.200 ElevenLabs Flash voice (~1,500 characters) $0.075 Twilio outbound (4 min at $0.014) $0.056 GPT-5.4 mini (~50k input, ~500 output) $0.040 Deepgram Nova-3 transcription (4 min) $0.019 Total ~$0.39 (about 10 cents a minute) The model came in second cheapest. Half the total went to the platform fee, and the voice cost almost double what the LLM did. People always ask where 50k input tokens come from when the model only writes about 500. Most stacks send the whole conversation back to the model on every turn. Turn 1 is the system prompt plus one line from the caller, about 2,000 tokens. By turn 20 it’s carrying everything said so far, closer to 3,000. So cost doesn’t grow in a straight line with call length. Run the same pace for 10 minutes and the call is 2.5x longer, but it ends up with over 3x the input tokens. Tool results are where it gets ugly. Say the agent pulls the lead’s CRM record at turn 5 and gets back 1,500 tokens of JSON. That record rides along for the other 15 turns and adds about half again to the call’s input. If the agent only needs a name, a budget and a property type, pass those three and drop the rest. Every cost thread in here ends up as an argument about which model is cheapest per token, and it bugs me a bit. I get why, it’s the fun part to argue about. But the whole LLM line on this call is 4 cents, so a model at half the price saves you 2 cents a call while the platform fee sits at 20. If you want the LLM line smaller anyway, turn on prompt caching before you go model shopping. Most of every turn is the same system prompt and the same history, and cached input bills at about a tenth of the normal price. On this call that takes the LLM from 4 cents to under 1. Model choice does start to matter once you’re on a big model. On GPT-5.5 with no caching, this same call’s LLM line is about 27 cents which makes it the biggest thing on the receipt. Caching drops it to about 5. I’d put the effort into the other lines. Voice is billed per character, so shorter agent turns cut it directly. They also stop callers from talking over the agent which they do constantly when it rambles. I care more about cost per real conversation than cost per call. If half your dials are 30 second voicemails, your real number is a lot higher than the receipt above. TLDR: About 39 cents for a 4 minute call at list prices, half of it platform fee and only 10% LLM. Because every turn resends the history, input tokens grow faster than call length, so turn on caching and trim tool results before you go looking for a cheaper model.