Short answer
AI models are priced per token, a small piece of text averaging about three quarters of an English word. You pay for the tokens you send (instructions, your information, the conversation) and for the tokens the model writes back, which usually cost several times more. The choice of model changes the bill dramatically: on Anthropic's price list in October 2026, the same customer reply costs about twenty times more on a mid-tier model than on the smallest one.
01AI is billed like mobile data
Mobile data is billed by the megabyte, whatever you do with it. AI is billed by the token. A token is a small piece of text. In English it averages about three quarters of a word, so 100 tokens is roughly 75 words. Punctuation, numbers and spaces count too.
You pay in two directions:
- Input tokens: everything you send the model. Your instructions, the business information it needs, the conversation so far, any documents and tool results.
- Output tokens: everything the model writes back. These usually cost several times more per token than input.
02What a price list looks like
A slice of Anthropic's published API prices for its current Claude models, in US dollars per million tokens, as listed on 11 October 2026. Other providers publish similar lists.
| Model | Input, per million tokens | Output, per million tokens | Typical use |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | High-volume, simpler work: sorting messages, short replies, pulling out details |
| Claude Sonnet 5.5 | $2 | $10 | Most business work: good replies, reasoning, agents |
| Claude Opus 5.5 | $4 | $20 | The hardest reasoning and long, complex tasks |
- Model
- Claude Haiku 5.5
- Input, per million tokens
- $0.10
- Output, per million tokens
- $0.50
- Typical use
- High-volume, simpler work: sorting messages, short replies, pulling out details
- Model
- Claude Sonnet 5.5
- Input, per million tokens
- $2
- Output, per million tokens
- $10
- Typical use
- Most business work: good replies, reasoning, agents
- Model
- Claude Opus 5.5
- Input, per million tokens
- $4
- Output, per million tokens
- $20
- Typical use
- The hardest reasoning and long, complex tasks
Prices change, usually downward, so check the provider's current page before budgeting. Haiku 5.5's listed price applies to prompts of up to 100,000 tokens.
03A worked example: AI replies to customers
Say each customer reply needs about 2,000 input tokens (instructions, product and delivery information, the conversation so far) and produces about 300 output tokens, and the business handles 1,000 such replies a day.
| Model | Cost per reply | Per day (1,000 replies) | Per month (30 days) |
|---|---|---|---|
| Claude Haiku 5.5 | $0.00035 | $0.35 | about $10.50 |
| Claude Sonnet 5.5 | $0.007 | $7 | about $210 |
| Claude Opus 5.5 | $0.014 | $14 | about $420 |
- Model
- Claude Haiku 5.5
- Cost per reply
- $0.00035
- Per day (1,000 replies)
- $0.35
- Per month (30 days)
- about $10.50
- Model
- Claude Sonnet 5.5
- Cost per reply
- $0.007
- Per day (1,000 replies)
- $7
- Per month (30 days)
- about $210
- Model
- Claude Opus 5.5
- Cost per reply
- $0.014
- Per day (1,000 replies)
- $14
- Per month (30 days)
- about $420
The same job shows a forty-fold difference between the cheapest and the most capable model. The right answer is rarely “always the best model”. It is usually a small, fast model for the simple majority of messages and a stronger one for the hard cases.
04Why agents cost more than chat
An agent does not send one request. It reads, calls a tool, reads the result, thinks and calls another tool, and each step sends the growing context again.
Take an agent that works through ten steps, each reading about 5,000 tokens and writing 500. That is 50,000 input and 5,000 output tokens, about $0.15 per task on Sonnet 5.5, or roughly twenty times a single reply. Cheap for a task that saves a person fifteen minutes. Expensive if it runs on every message for no reason.
05Six ways to keep the bill under control
- Use the smallest model that does the job well, and send only the hard cases to bigger models.
- Keep instructions and context lean. Every unnecessary paragraph is paid for on every request.
- Use prompt caching. Providers charge far less for repeated, unchanged parts of a request, such as fixed instructions and the product list. On Anthropic's list, cached Sonnet 5.5 input costs $0.10 per million tokens instead of $2.
- Send only the relevant information, the way RAG does, rather than everything you have.
- Use discounted batch processing for work that does not need an instant answer, such as overnight reports.
- Set daily budgets and alerts per feature, and track the cost per resolved conversation, not only the total.
06The full cost is more than tokens
Model usage is often the smallest line. A realistic monthly budget for an AI feature also includes:
- Channel fees, such as WhatsApp template messages charged by Meta.
- Hosting, databases and search for your information.
- Monitoring, logs and storage of conversations.
- People's time to review samples and handle the cases the AI passes on.
- Maintenance as models, prices and your products change.
Per-person subscriptions, such as business plans for ChatGPT or Claude, suit staff who use AI directly. Customer-facing systems and automations use the API and are priced by tokens. Many businesses need both.
Questions
What is a token in AI?
A token is a small piece of text that a model reads or writes, often part of a word. In English, 100 tokens is roughly 75 words. AI usage is measured and priced in tokens.
Why does Bangla text cost more?
Most tokenizers split Bangla into more pieces than English, so the same message uses more tokens. The difference varies by model, so test with your own messages.
Is a ChatGPT or Claude subscription enough for a business?
For staff using AI to write and research, a per-person plan can be enough. For AI that answers customers, connects to your systems or runs on its own, you need API access, priced by usage.
How can I estimate my AI cost before building?
Estimate requests per day, measure input and output tokens per request with a few real examples, multiply by the provider's current prices, then add 30 to 50 percent for retries, longer conversations and growth.
Last updated