AI token cost calculator for spreadsheets: what pasting your sheet costs per question

By Killian, founder · Updated 28 October 2026 · Markdown

A pasted spreadsheet costs roughly 250,000 tokens per megabyte of CSV (about 2.6 tokens per cell on a typical orders sheet), and you pay for all of it again on every question. Our 50,000-row, 8-column orders file is 4.1 MB: about 1,030,000 tokens, more than a 1M-token context window holds, and $4.12 of input per question at Claude Opus 5.5’s price of $4 per million input tokens. Asking the same question through datatomcp took about 2,200 tokens, under one cent. Enter your own sheet’s size and pick your model in the calculator above.

How the calculator works

1. Characters. Rows × columns × characters per cell gives the size of the sheet as text. The default of 10.3 characters per cell (commas included) is measured: our orders file has 400,000 cells and is 4,118,185 bytes. Your sheet’s number is its CSV file size in bytes divided by rows × columns.
2. Tokens. We count 4 bytes of CSV per token. When the founder pasted the same file into Claude, Claude reported the attachment at about 975,000 tokens, which is 4.2 bytes per token, so the estimate is slightly on the high side.
3. Cost. Tokens × your model’s input price. Picking a model fills in its list price, including the higher price some vendors charge for long prompts; the LLM API pricing comparison lists every model with its source. Output tokens are left out: the answer is the same length either way.

Sizes at a glance

8 columns, 10.3 characters per cell, $4 per million input tokens, 1M-token context window:

Why you pay again on every question

A chat model has no memory between requests. Each question sends the whole conversation again, the pasted sheet included, so ten questions read the sheet ten times. Prompt caching can make those repeat reads cheaper, but the sheet still fills the same share of the context window, leaving less room for the question, the answer and anything else you paste.

On a chat subscription you don’t see a bill per token; the cost shows up as limits instead. When the founder pasted the 50,000-row file into Claude on the free plan, it refused before answering: “This request exceeds Claude’s context limit by about 1,252,640 tokens.” On a 5,000-row file Claude answered correctly, because it ran code on the attachment. Small sheets are fine; the trouble starts when the sheet is bigger than the window.

The other way: ask, don’t paste

Connect the sheet instead and the model sends a question, not the data. datatomcp keeps the rows in a database and returns only the result: for “How many orders came from France in September 2026, and what was the total revenue?”, Claude listed the datasets, read the column descriptions and ran one aggregate, about 8,700 bytes in all, and got 277 orders and €50,279.50, the exact answer. That cost does not grow with the number of rows; the column descriptions grow a little with wider sheets.

1. Import the sheet. Sign in at app.datatomcp.com and upload a CSV or pick a Google Sheet (100,000 rows free, 2 million on Pro).
2. Connect your AI. Add https://app.datatomcp.com/mcp as a connector in Claude or another MCP client.
3. Ask normally. Totals and breakdowns are computed by the database, so the answer is counted, not read off pasted text.

FAQ

Is this the same for ChatGPT and Gemini? Token counts differ a little between tokenizers, and prices differ a lot. Pick the model in the calculator, or type its input price; the token estimate stays close. What a spreadsheet question costs on each model has the full table.

Why characters per cell and not file size? Both work. If you know the file size, 1 MB of CSV is about 250,000 tokens. Rows and columns are easier to read off a sheet you have open.

Does uploading the file instead of pasting change the cost? It depends on the app. If the whole file is read into the conversation, it costs the same as pasting: Claude counted our 50,000-row attachment at about 975,000 tokens. Some apps run code on the file instead, which is how Claude got the 5,000-row file right.