A worked estimate you can check line by line. Change the numbers to your own in the cost calculator.
What we assume
Each ticket takes 4 replies. Every reply sends a 3,000-token system prompt with tool definitions (cacheable) plus about 4,000 new tokens of retrieved help-center articles, account data and the customer's message, and gets a 350-token answer.
Cost per ticket
| Model | No caching | With caching | Per 1,000 tickets, with caching |
|---|---|---|---|
| Claude Fable 5.2 | $0.280 | $0.186 | $186 |
| Claude Opus 5.5 | $0.140 | $0.094 | $94.40 |
| Claude Sonnet 5.5 | $0.070 | $0.047 | $47.20 |
| Claude Haiku 5.5 | $0.0035 | $0.0024 | $2.42 |
Cost per month
| Model | Small team (500 tickets/month) | Busy help desk (10,000 tickets/month) |
|---|---|---|
| Fable 5.2 | $93.20 | $1,864 |
| Opus 5.5 | $47.20 | $944 |
| Sonnet 5.5 | $23.60 | $472 |
| Haiku 5.5 | $1.21 | $24.20 |
Where the money goes
- Retrieved documents are usually the biggest part of the bill. Retrieving fewer, better chunks saves more than switching models.
- Tool definitions and the system prompt repeat on every reply, so they are the part prompt caching makes cheap.
- Compare the cost per ticket with what a ticket costs your team today; that number decides which model you can afford.
To check what a plan price leaves you after these costs, use the AI SaaS margin calculator. To see how much caching saves on your own prompt sizes, use the prompt caching savings calculator.
Questions
How much does one ticket cost with the Claude API?
With prompt caching, from about $0.0024 on Claude Haiku 5.5 to about $0.186 on Claude Fable 5.2, for a ticket of 4 calls × (3,000 reused + 4,000 new input tokens, 350 output tokens). Without caching it is $0.0035 to $0.280.
How much does it cost per month?
Small team (500 tickets/month): about $1.21 on Haiku 5.5 up to $93.20 on Fable 5.2. Busy help desk (10,000 tickets/month): about $24.20 on Haiku 5.5 up to $1,864 on Fable 5.2, with caching.
How can I make it cheaper?
Retrieved documents are usually the biggest part of the bill. Retrieving fewer, better chunks saves more than switching models.