FREE TOOL
What will your AI chatbot cost to run?
Describe how your customers chat and see the API bill per conversation and per month, model by model. It counts the chat history that is sent again with every message, which quick estimates leave out.
| Model | Per conversation | Per month | List price per 1M tokens |
|---|---|---|---|
| GPT-5 nanoOpenAI | $0.00088 | $0.88 | $0.05 in, $0.4 out |
| GPT-OSS 20BGroq | $0.0012 | $1.18 | $0.075 in, $0.3 out |
| Gemini 2.5 Flash-LiteGoogle | $0.0016 | $1.57 | $0.1 in, $0.4 out |
| GPT-4o miniOpenAI | $0.0024 | $2.35 | $0.15 in, $0.6 out |
| GPT-5 miniOpenAI | $0.0044 | $4.40 | $0.25 in, $2 out |
| Gemini 2.5 FlashGoogle | $0.0053 | $5.33 | $0.3 in, $2.5 out |
| GPT-4.1 miniOpenAI | $0.0063 | $6.27 | $0.4 in, $1.6 out |
| Gemini 3 FlashGoogle | $0.0083 | $8.32 | $0.5 in, $3 out |
| Claude Haiku 4.5Anthropic | $0.016 | $16.17 | $1 in, $5 out |
| GPT-5OpenAI | $0.022 | $22.01 | $1.25 in, $10 out |
| Gemini 2.5 ProGoogle | $0.022 | $22.01 | $1.25 in, $10 out |
| Claude Sonnet 5.5Anthropic | $0.032 | $32.33 | $2 in, $10 out |
| Claude Opus 5.5Anthropic | $0.065 | $64.67 | $4 in, $20 out |
Estimates from public list prices in US dollars, last checked 2026-10-13. One token is counted as three quarters of an English word. Cached and batch requests are billed lower, so a real bill is usually below these figures.
THE PART PEOPLE MISS
Why a long chat costs more per message.
A chat model has no memory. Each time a customer writes, the app sends the instructions, the knowledge-base text and the whole chat so far, and you pay for all of it again. A chat twice as long sends about four times as much history.
| Customer messages | Input tokens billed | Average per message |
|---|---|---|
| 1 | 2,027 | 2,027 |
| 5 | 11,205 | 2,241 |
| 10 | 25,085 | 2,509 |
| 20 | 60,870 | 3,044 |
Worked out from the numbers you entered above.
GOOD TO KNOW
How the numbers are worked out.
How is the cost of a chatbot conversation calculated?
Providers charge per token, with one price for the text you send (input) and a higher one for the text the model writes (output). The calculator adds up every token sent and written across the whole conversation and multiplies each side by the model's list price.
Why does the history get sent again?
The model keeps nothing between messages. To answer the fifth question in context, the app sends the first four questions and answers along with it. That is why input tokens, not replies, make up most of a chatbot bill.
How many tokens is a word?
In English one token is about three quarters of a word, so 100 words are roughly 133 tokens. Other languages, numbers and code use more tokens per word, so add a margin if your customers do not write in English.
What is the knowledge-base text?
An assistant that answers from your own documents looks up the passages that match each question and sends them to the model with it. Those passages are usually the largest part of every request.
Are these prices exact?
They are the providers' public list prices on the date shown under the table, and they change. Discounts for cached or batch requests are not counted, so treat the result as an upper estimate and check the provider's pricing page before you budget.
Can I run a chatbot with no API bill?
Yes, with a model you host yourself, for example through Ollama. There is no charge per token, but you pay for the computer it runs on, and small local models answer less reliably than the hosted ones.
Does Crescelo add a markup on AI usage?
No. You connect your own key from OpenAI, Anthropic, Google, Groq or DeepSeek, or a self-hosted model, and the provider bills you directly. The Crescelo plan price is separate.
Pay the provider, not a markup.
Crescelo connects to your own AI key, so the bill above goes straight to the provider at their price. We add nothing on top. See how to connect a key