Skip to content

Free AI tool · run cost per month

AI agent and RAG cost calculator.

Describe your agent: how many tasks a day, how many model calls each takes, the tokens per call, how much of the prompt is cached and how often calls are retried. Add the retrieval side: documents to embed, chunking, re-indexing, queries and reranking. The calculator prices it from the same dated list prices as our LLM pricing table and ranks what drives the bill.

The agent

Each reasoning or tool step is one call.

Instructions, history and tool results.

Billed as output.

The repeated prefix: system prompt, tools, history.

Calls repeated after errors or failed checks.

Retrieval (RAG)
Re-index

Documents that changed since the last run.

Reranking

Monthly run cost

$550.63

Claude Sonnet 5.5, 36,500 model calls a month, 25,000 chunks indexed. Prices checked on 18 October 2026.

Per task
$0.091
Per day
$18.10
Per year
$6,607.50
Indexing, one-off
$0.220
11,000,000 tokens embedded

What drives the cost

  • Reasoning tokens$219.00 · 40%
  • Model output$146.00 · 27%
  • Model input (not cached)$116.80 · 21%
  • Retrieved context sent to the model$30.42 · 6%
  • Retries$24.53 · 4%
  • Model input (cached)$8.76 · 2%
  • Reranking$5.02 · 0.91%
  • Re-indexing (embeddings)$0.096 · 0.02%
  • Query embeddings$0.0061 · 0.00%

Model calls are priced like the LLM pricing table: uncached and cached input, output and reasoning tokens at the model's per-1M prices, plus the same again for the retry rate. Retrieved chunks are added to the model's input once per query. Reranking uses Voyage's formula: query tokens × chunks reranked, plus the tokens of those chunks. Vector database hosting, tool compute and cache write fees are not included.

Planning an AI agent? Talk to us about building it, with the cost per task designed in from the start.

AI developmentTalk to us

Prices checked on 18 October 2026. Public list prices in US dollars per 1M tokens, copied by script from models.dev (MIT, list prices, cached-input prices, context windows) and LiteLLM model price file (MIT, batch api prices and reranker prices). Nothing is typed by hand. Confirm a price on the provider's own page before you commit to it; providers change prices and add surcharges (long prompts, regions, priority tiers) that a summary table cannot show.

Rules applied, and where we read them

Where the money goes

Agents are billed per call, and calls multiply.

  • 01

    Steps multiply every token

    An agent that plans, calls a tool and checks its work makes several model calls per task, and each call resends the instructions and the history so far. Input tokens per call times calls per task is often the largest line.

  • 02

    Caching is a large lever

    The system prompt, tool definitions and earlier turns repeat across calls. When the provider serves them from its prompt cache they are billed at the cached price, so the cache hit rate can move the monthly total a long way. Try it in the calculator.

  • 03

    RAG costs little to index, more to read

    Embedding a document set is often a small one-off cost. The recurring cost is the retrieved text sent to the model with every query, so the chunk size and the number of chunks you pass in often matter more than the embedding price.

FAQs · AI agent cost

AI agent and RAG costs · frequently asked questions

How the estimate is built, what it leaves out, and how to make it match your agent.

It depends on four numbers more than anything else: tasks per day, model calls per task, tokens per call and the model’s price. Enter yours and the calculator multiplies them out per month, with caching, retries and retrieval added as separate lines so you can see which one drives the bill.
Run the agent on a few real tasks and read the usage the API returns with each response: input, cached input, output and reasoning tokens. Average them per call. Our token counter gives a quick count of a system prompt or a document if you have no logs yet.
Documents are split into chunks with the overlap you set and embedded once, which is the one-off indexing cost. Each month adds the share you re-embed, the embedding of every query, the retrieved chunks sent to the model as input, and reranking if you use it, priced with Voyage’s published formula.
Hosting for the vector database, the compute behind the tools your agent calls, web search and tool fees, cache write and storage fees, and the engineering time to build and maintain the agent. The figure is the model and embedding bill only.
From the same snapshot as our LLM API pricing table: public list prices copied by script from the open models.dev and LiteLLM price files, with the date they were checked printed on the page. Reranker prices come from LiteLLM and match Voyage’s own pricing page.

Next step

Planning an AI agent? Talk to us.

QalbIT designs and builds AI agents and retrieval into business software, with the model, caching and retrieval chosen against a cost per task you agree up front.

AI development

Keep exploring

More free tools

More free tools

Each one opens on a worked example and shows its answer without a signup.

All free tools