What does this voice AI LLM cost calculator actually estimate?

It estimates the model cost of running an AI voice agent: the dollars you pay the large language model provider for each call and across a month, given your system prompt size, average call length, and call volume. It does not include speech-to-text, text-to-speech, or telephony, which are separate line items in a voice stack. The point is to isolate the one cost most operators get wrong, the LLM, and show how much it changes when you swap models. A long system prompt resent on every conversational turn is the silent budget killer, and this tool makes that visible. Pick a model, set your prompt size and call length, and you get a per-call number you can multiply against real volume before you ever wire up a provider key.

How is the per-call cost calculated?

For text pipeline models the tool models a real conversation. It assumes roughly three back-and-forth turns per minute. On every turn the agent sends the system prompt plus the running transcript so far, which grows as the call goes on, and the model returns a response. It sums input tokens across all turns, adds output tokens, and prices each against the model's published input and output rates per million tokens. If you turn on prompt caching, the repeated system prompt is billed at the much lower cached-read rate, which is how most production voice agents are actually run. For GPT Realtime, a speech-to-speech model, cost is driven by audio tokens instead, so the tool prices it per minute of audio in and out. All numbers are indicative; confirm current rates with each provider.

Why does my system prompt size matter so much?

Because in a voice agent the system prompt is not sent once, it is sent on every single turn of the conversation. A 3,000-token prompt on a five-minute call with fifteen turns means you are paying to process those 3,000 tokens fifteen times, before you count the growing transcript. Double the prompt and you roughly double the input cost. This is why a bloated prompt full of edge cases and examples can quietly cost more than the model choice itself. Prompt caching helps a lot, because the static portion of the prompt is billed at a fraction of the normal input rate after the first turn. The practical takeaway is to keep the system prompt tight, push reference material into a knowledge base instead of the prompt, and turn on caching wherever the model supports it.

Is GPT Realtime cheaper than a text pipeline?

Usually not on raw model cost, but it can be cheaper overall. Speech-to-speech models like GPT Realtime price on audio tokens, which run far higher per minute than text tokens, so the LLM line item is typically larger. What you save is the rest of the pipeline: you no longer pay separately for speech-to-text and text-to-speech, latency drops because there is no transcription hop, and the engineering to stitch three services together disappears. So the right comparison is not LLM versus LLM, it is total cost and quality of the whole stack. This calculator shows the model cost clearly so you can weigh it against the pipeline you would otherwise build. For most agencies the deciding factor ends up being latency and naturalness, not the per-minute model fee.

How accurate are these numbers for budgeting?

Treat them as a planning estimate, not a bill. The token math is based on reasonable assumptions about conversation density and message length, and the prices are the published list rates at the time of writing, which providers change. Real calls vary: some callers are terse, some ramble, some agents use tools and function calls that add tokens. The value of the tool is relative, not absolute. It is very good at telling you that switching from one model to another roughly halves your cost, or that trimming your prompt by a thousand tokens saves a few hundred dollars a month at your volume. Use it to make decisions and to sanity-check a provider quote, then validate against a few real call logs before you commit to a model at scale.

How does Hermes change the LLM cost math for my agency?

Hermes is the platform agencies run their voice agents on, white-labeled end to end, so instead of buying model credits, speech services, telephony, a CRM, and a developer separately, you get one platform with usage billed transparently. You build the agent, connect a number, and launch inbound and outbound campaigns under your own brand, and your clients never see the word Hermes. Pricing is straightforward: Starter is $149 a month with three white-label workspaces and 300 included minutes, Business is $399 with 1,000 minutes, and Agency is $699 with 2,000 minutes, all at $0.24 per overage minute. Because the model, voice, and telephony costs are bundled into a known per-minute rate, you can quote clients a fixed price and keep your margin predictable. Want the full breakdown for your agency? Book 20 minutes and we will map your real per-call cost and your first agent launch with you live.

free tool

Voice AI LLM cost calculator.

See what the model actually costs per call and per month for your AI voice agent. Compare GPT-4o, GPT Realtime, Claude, and Gemini, with prompt caching on or off. Built for operators sizing a stack. No email, no signup.

/ your voice agent

System prompt size (tokens)Resent on every turn. ~750 words = 1,000 tokensAverage call length (minutes)~3 conversational turns per minuteCalls per monthAcross all clients on this agentModelText pipeline (STT → LLM → TTS). Balanced quality and cost.

Prompt caching on (system prompt billed at the cached-read rate after the first turn)

/ cost per call

$0.0431

GPT-4o, ~12 turns

/ cost per month

$43.05

1,000 calls

/ cheapest model

Gemini 2.0 Flash

$41.78/mo less than your pick

/ model comparison (cheapest first)

Gemini 2.0 Flashtext pipeline · $0.0013/call

$1.27/mo

GPT-4o minitext pipeline · $0.0026/call

$2.58/mo

Claude Sonnettext pipeline · $0.0320/call

$32.04/mo

GPT-4oyour picktext pipeline · $0.0431/call

$43.05/mo

GPT Realtimespeech-to-speech · $0.4800/call

$480/mo

LLM cost only. STT, TTS, and telephony are billed separately. Indicative list pricing — verify with each provider.

Book a 20-min walkthrough to cut this number.

We will trim your prompt, pick the right model, and bundle model, voice, and telephony into one transparent per-minute rate. One platform, your brand, from $149/month.

Book your walkthrough →

Your inputs are encoded in the URL. Send it to a partner and they see the same numbers.

how it works

The system prompt is the line item everyone underprices.

When operators price a voice agent, they think about the per-minute rate they see on a provider page and stop there. The trap is that a voice agent does not send its prompt once. On every conversational turn it resends the full system prompt plus the running transcript, then the model replies. A five-minute call can be fifteen turns, which means a long, example-stuffed prompt gets processed fifteen times in a single call. That is why two agencies running the same model can have wildly different bills: one wrote a tight prompt and pushed reference material into a knowledge base, the other pasted everything into the system prompt.

This calculator makes that visible. It models roughly three turns per minute, grows the transcript turn by turn, and prices input and output tokens against each model's published rates. Toggle prompt caching and watch the repeated prompt drop to the cached-read rate, which is how production agents are actually run. Switch between a text pipeline model and GPT Realtime, a speech-to-speech model priced on audio tokens, and you can see the real trade: realtime usually costs more per minute on the model but removes the separate speech-to-text and text-to-speech bills and cuts latency.

Use the per-call number to sanity-check a provider quote, then multiply by real volume to see the monthly line. The numbers are indicative and providers change prices, so validate against a few real call logs before you lock in a model. When you are ready to put the agent behind a real number under your own brand, that is where Hermes comes in: model, voice, and telephony bundled into one transparent per-minute rate starting at $149 a month, so you can quote clients a fixed price and keep your margin predictable.

frequently asked

Common questions.

next step

Know your model cost? Run the whole agent on Hermes.

One platform. Your brand. From $149/month, with model, voice, and telephony bundled into one transparent per-minute rate.

Book a 20-min walkthrough Apply for the Founders' Beta

Want the full breakdown for your agency? Book 20 minutes and we will map your real per-call cost, the right plan, and your first agent launch with you live.

Voice AI LLM cost calculator.

The system prompt is the line item everyone underprices.

Common questions.

Know your model cost? Run the whole agent on Hermes.

/ Keep reading