LLM API pricing comparison: Claude, GPT and Gemini per million tokens

By Killian, founder · Updated 28 October 2026 · Markdown

Current LLM API prices run from $0.10 per million input tokens (Claude Haiku 5.5 on prompts up to 100K tokens, GPT-6 Luna) to $10 (Claude Fable 5.1, GPT-6 Astra), and output costs five times the input price on every model below except Gemini 3.1 Flash-Lite, where it is six. All eleven models have a context window of about one million tokens. For a spreadsheet question, the model matters less than how much of the sheet you send: pasting a 5,000-row, 8-column sheet costs $0.01 per question on GPT-6 Luna and $1.03 on Claude Fable 5.1, and asking the same question through datatomcp sends about 48 times less data on any of them. Change the rows and columns in the table above to price your own sheet.

LLM API prices, per million tokens

Standard pay-as-you-go text prices in US dollars, checked on 28 October 2026 on each vendor’s pricing page. Batch, cached-input and priority prices are left out.

Context windows come from the vendors’ model pages: Claude models overview (1,000,000 tokens), OpenAI compare models (1,050,000) and the Gemini model pages, e.g. Gemini 4 Argon (1,048,576). Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon share the same $2 / $10 price until Gemini 4 Argon doubles on 1 December.

What one spreadsheet question costs on each model

Input cost of one question, 8 columns, 10.3 characters per cell (measured on our orders file):

* Billed at the model’s long-prompt price: the paste is over 100K tokens (Haiku) or 272K tokens (OpenAI).

The pasted sheets are about 20,600, 103,000 and 412,000 tokens. The datatomcp column is about 2,200 tokens, whatever the size of the sheet: the tool list, the dataset’s column descriptions and the answer, measured on a real question. At 50,000 rows the paste is about 1,030,000 tokens, right at the edge of every window here, before the question and the answer are added. When the founder pasted that file into Claude (free plan), it refused before answering.

Cheaper model, or less data?

There are two ways to cut the bill for a spreadsheet question. Switching from Claude Opus 5.5 to Claude Haiku 5.5 takes a pasted 5,000-row sheet from $0.41 to $0.05 a question. Sending less data takes the same question to $0.0087 on Opus itself, because the model reads about 8.7 KB of tool calls and results instead of the 412 KB sheet.

datatomcp does not make any model cheaper. The saving comes only from sending less: the rows stay in a database, the model asks for a total or a filtered list, and the database sends back the result. That is also why it keeps working when the sheet outgrows the window. On our 50,000-row file, Claude’s free plan refused the paste, and through datatomcp it answered “277 orders, €50,279.50”, the exact figure.

To try it, connect a CSV or Google Sheet (Claude setup, any other MCP client) and ask the same question you would have pasted the sheet for.

What the list price leaves out

• Long-prompt tiers. OpenAI bills the whole request at the higher price once the prompt passes 272K input tokens, and Claude Haiku 5.5 does the same past 100K. A pasted sheet is exactly the kind of prompt that crosses the line. Claude Fable, Opus and Sonnet charge the same per token across the full 1M window.
• Announced changes. Google’s pricing page lists Gemini 4 Argon at $4 / $20 from 1 December 2026 and Gemini 3.8 Flash at $1.50 / $7.50 from 1 January 2027.
• Caching and batch. Anthropic and OpenAI both sell cached input at a fraction of the standard price (GPT-6.1 Sol: $0.10 instead of $2), which makes repeat reads of the same pasted sheet cheaper. The sheet still takes the same share of the context window.
• Tokenizers. Each vendor counts tokens differently, so the same CSV is a slightly different number of tokens on each. We use 4 bytes per token for all of them; Claude counted our 4.1 MB file at about 975,000 tokens, 4.2 bytes per token.

How the costs are worked out

Tokens per paste are rows × columns × characters per cell ÷ 4, the same estimate as the spreadsheet token cost calculator, which also counts repeat questions and the share of the window used. Cost is tokens × the model’s input price for a prompt of that size. Output is left out of the per-question cost: the answer is about the same length whether the sheet was pasted or queried. The datatomcp figure is one measured session on our 50,000-row, 8-column test file: 8,663 bytes in total.

FAQ

Which LLM API is cheapest? On input price, Claude Haiku 5.5 (prompts up to 100K tokens) and GPT-6 Luna tie at $0.10 per million input tokens and $0.50 per million output. For prompts over 100K tokens, GPT-6 Luna stays at $0.10 up to 272K, while Haiku moves to $0.50.

Is Claude cheaper than GPT? At the same price level they match: Claude Sonnet 5.5 and GPT-6.1 Sol are both $2 / $10, Claude Fable 5.1 and GPT-6 Astra both $10 / $50. The difference is long prompts: past 272K input tokens the OpenAI models cost more, while Claude Fable, Opus and Sonnet keep the same price up to 1M.

How many tokens is a spreadsheet? About 250,000 tokens per megabyte of CSV, or roughly 2.6 tokens per cell on a typical orders sheet. A 5,000-row, 8-column sheet is about 103,000 tokens.

Do these prices apply to ChatGPT, Claude or Gemini subscriptions? No. They are API prices, billed per token. On a chat subscription the same tokens show up as usage limits and context limits instead of a bill.