LLM cost calculator
Say what 1 run of your job reads and writes and how often it runs. See what it costs on every current model, a run, a day, a month or a year, or start from your budget and see which models fit it.
Prices read from each maker's page on 14 October 2026
The calculator
A short question and a few paragraphs back.
about 230 words
about 380 words
Morecache, thinking, several turns, batch
Text sent again unchanged costs the cached price.
For a task of several turns, such as an agent. Only this part has to fit in the window.
The answer
Cheapest: Qwen3.8-Flash at 84 cents a month. Dearest: Claude Fable 5.1 at $84.00, 100 times as much.
The bill, cheapest first · a month
- Input
- Cached input
- Output
- Thinking
Qwen3.8-FlashAlibaba
84 cents
0.028 cents a run
GPT-6 LunaOpenAI
84 cents
0.028 cents a run
Claude Haiku 5.5Anthropic
84 cents
0.028 cents a run
DeepSeek V4.1 FlashDeepSeek
$2.07
0.069 cents a run
Gemini 3.5 Flash-LiteGoogle
$4.02
0.13 cents a run
Gemini 3.8 FlashGoogle
$6.30
0.21 cents a run
DeepSeek V4 ProDeepSeek
$7.13
0.24 cents a run
Grok 4.7xAI
$10.80
0.36 cents a run
Qwen3.8-MaxAlibaba
$10.80
0.36 cents a run
GPT-6.1 SolOpenAI
$16.80
0.56 cents a run
List prices, no discounts. A month is 30 days. Where a maker publishes no cached price, cached input is charged as input.
What common jobs cost a month
Each job at the runs a day it starts with, priced on the cheapest and the dearest of the newest models whose window holds it. The gap is the point: choosing the model moves the bill more than anything else you do.
A chat question
100 runs a day
- From
- 84 cents on Qwen3.8-Flash
- Up to
- $84.00 on Claude Fable 5.1
A hard question, with thinking
100 runs a day
- From
- $2.11 on Qwen3.8-Flash
- Up to
- $219 on Claude Fable 5.1
A long report, summarised
20 runs a day
- From
- $4.20 on GPT-6 Luna
- Up to
- $420 on Claude Fable 5.1
A 40-turn agent task
10 runs a day
- From
- $21.96 on GPT-6 Luna
- Up to
- $2,196 on GPT-6 Astra
A support bot's reply
2,000 runs a day
- From
- $22.20 on GPT-6 Luna
- Up to
- $2,220 on GPT-6 Astra
Sort a message into a category
50,000 runs a day
- From
- $41.25 on GPT-6 Luna
- Up to
- $4,125 on Claude Fable 5.1
Read an invoice into fields
1,000 runs a day
- From
- $8.25 on GPT-6 Luna
- Up to
- $825 on Claude Fable 5.1
Review a pull request
50 runs a day
- From
- $7.13 on GPT-6 Luna
- Up to
- $713 on Claude Fable 5.1
How the sum works
A model is paid by the token, and every maker prints its prices for 1 million of them. One run of a job has 4 parts, each with its own price: the input the model reads for the first time, the input it reads again from its cache, the answer it writes, and the thinking it does before the answer. Thinking is billed at the output price even though you never read it.
The calculator multiplies each part by the model's price, adds them up for 1 run, then multiplies by the runs a day or a month. A month is 30 days. With "Answers can wait" it uses the batch price on the models that publish one.
A model whose window is too small for the largest prompt you send at once is greyed out, not hidden, so you can see what it would have cost. For a task of several turns, such as an agent, that largest prompt is far smaller than the total read across the turns.
It leaves out the first write to a cache (dearer than input on some models), tools such as web search, pictures and sound, and any discount agreed with a maker. If a figure here and the maker's page disagree, the maker's page is right.
Questions
- How do I work out what an LLM API will cost me?
- Count the tokens 1 run of your job reads and writes, multiply each by its price per million tokens, then multiply by the runs. Input, cached input, output and thinking each have their own price, and thinking is billed as output. The calculator does the sums for every current model at once.
- How many tokens is a word or a page?
- In English a token is about three quarters of a word, so 1,000 words is about 1,330 tokens and a page of 500 words about 670. Code and other languages take more tokens per word. Switch the calculator to words or pages and it converts for you.
- Why does the same job cost so much more on some models?
- Prices per token differ by a hundred times or more between the cheapest and the dearest model, and writing costs several times as much as reading. A short chat question asked 100 times a day costs 84 cents a month on Qwen3.8-Flash and $84.00 on Claude Fable 5.1.
- What is a batch price?
- A lower price, often half, for requests that can wait: you send them together and collect the answers later, within a day. Tick "Answers can wait" and the calculator uses it on every model that has one.
- Does a cache really lower the bill?
- Yes, when the model reads the same text again: a long system prompt, a document you ask about several times, or an agent sending its conversation on each turn. Cached input costs a fraction of the input price. Set the share of input read from cache with the slider.