What does it really cost to run AI?
Model APIs quote a token price. Production bills include speech, telephony, vector databases, retries, evals and human review — LLM inference is often only 15–35% of the total. These calculators do the math on all of it. Free, no signup.
Why trust the numbers?
Every model price on this site is pulled daily, automatically from OpenRouter's public pricing API and rendered into the calculators as-is — no hand-maintained tables. Embedding endpoints aren't listed by that API, so they come from a second public source (LiteLLM's community price list), marked as such. Rates that aren't per-token (speech, telephony, vector DB) are editable inputs with the source and date stated next to them. Our method page shows every formula. If a number can't be sourced, it isn't shown.
→ See today's full model price table · → What changed in prices recently
Start with the whole bill, not the token price
Every provider publishes a per-token price and none of them publishes the bill. A production AI system pays for the services wrapped around the model, the data layer underneath it, and the operational overhead above it — so an estimate built from token pricing alone is not conservative, it is simply incomplete.
- Building a phone or voice product — start with the voice agent calculator. Four meters run per minute and the model is usually the smallest of them.
- Adding search or a knowledge assistant — start with RAG pipeline cost. The retrieval path, not the model, is where the money goes.
- Planning a support assistant — start with chatbot cost, then check what a human handoff does to the economics.
- Deciding whether to fine-tune — start with fine-tuning cost. It pays only when it shrinks a prompt or lets you downgrade a model.
- Generating images at volume — start with image API cost. Your regeneration rate matters more than your vendor's price list.
- Sizing a full project — start with AI project total cost. It shows the split, which is the part budgets usually get backwards.
The mistake almost every AI cost estimate makes
| Cost | On the vendor's pricing page? | Who bills it |
|---|---|---|
| LLM inference | Yes | Your model provider |
| Speech, telephony, image APIs | No — different vendors entirely | Two or three more suppliers |
| Vector database and retrieval storage | No | A hosting provider |
| Evals, tracing, observability | No | An observability vendor |
| Retries and fallbacks | No — invisible by design | Your model provider again |
| Human review and escalation | No | Your own payroll |
Published breakdowns of running production systems land on the same range repeatedly: LLM inference accounts for roughly 15–35% of total operating cost. That is why this site is organised around the stack rather than around the token price. It is also why the most common cost question — "which model is cheapest?" — so often turns out to be the wrong question.
What is on this site
| Page | Answers |
|---|---|
| AI project total cost | What does the whole thing cost per month, and what share is the model? |
| Voice agent cost | Cost per call and per minute across speech, model, synthesis and telephony. |
| AI SDR cost | Cost per lead and, more usefully, cost per meeting booked. |
| RAG pipeline cost | One-time indexing, recurring query cost, and the refresh loop. |
| Chatbot cost | Cost per conversation, and why length beats volume. |
| Fine-tuning cost | One-time training versus monthly serving, and whether it pays back. |
| Image API cost | Generation at scale, with regeneration overhead included. |
| Embedding cost | Indexing any corpus, and the query loop that never stops. |
| Model price table | Today's per-token prices for every model, with CSV download. |
| Price changes | What moved overnight, and by how much. |