AI feature cost-per-user calculator
Pick what the feature does, how many people use it, and the model. You get the model cost per use, per user per month, and for all users, with prompt caching and Batch API discounts, and a warning when a user costs too much of what they pay.
Result on Claude Sonnet 5.5
Per conversation
$0.0290
Per user / month
$0.58
Heavy user / month
$2.90
Bill for 1,000 users
$579.52
Model cost is 2.9% of the price for a typical user and 14.5% for a heavy one.
Tokens per conversation: 329 fresh input, 24,041 read from cache, 2,045 written to cache, 2,080 output.
Prompt size at each turn. The whole history is sent again every time. Dark is read from cache.
The same conversation on every model, cheapest first
| Model | Per conversation | Per user / month | Share of price | Bill for all users |
|---|---|---|---|---|
| $0.0012 | $0.0241 | 0.1% | $24.14 | |
| $0.0016 | $0.0314 | 0.2% | $31.38 | |
| $0.0027 | $0.0547 | 0.3% | $54.68 | |
| $0.0075 | $0.15 | 0.7% | $149.60 | |
| $0.0100 | $0.20 | 1.0% | $199.14 | |
| $0.0101 | $0.20 | 1.0% | $201.92 | |
| $0.0212 | $0.42 | 2.1% | $424.80 | |
| $0.0223 | $0.45 | 2.2% | $445.79 | |
| $0.0273 | $0.55 | 2.7% | $546.77 | |
| $0.0290 | $0.58 | 2.9% | $579.52 | |
| $0.0362 | $0.72 | 3.6% | $724.16 | |
| $0.0425 | $0.85 | 4.2% | $849.60 | |
| $0.0483 | $0.97 | 4.8% | $965.55 | |
| $0.0566 | $1.13 | 5.7% | $1,132.80 | |
| $0.0580 | $1.16 | 5.8% | $1,159.05 | |
| $0.0598 | $1.20 | 6.0% | $1,196.80 | |
| $0.11 | $2.27 | 11.3% | $2,265.60 | |
| $0.12 | $2.41 | 12.1% | $2,413.87 | |
| $0.14 | $2.78 | 13.9% | $2,777.41 |
Caching assumes the shared instructions are already warm from other users and each turn follows the last within the cache lifetime. Prompts shorter than the provider's minimum are not cached: 512 tokens on Claude 5.5 models, 1,024 on OpenAI, 4,096 on Gemini. Model cost only. Hosting, Stripe fees, retries, and web search calls are extra.
Cost per user per month, four example features
The calculator's starting values, run on seven models with prompt caching on and no Batch API. The inputs are our examples, not measurements from a client product.
- Summarise a document: A 2,000-word document (about 4 pages) in, a 200-word summary out, 30 a month.
- Extract fields from a document: Instructions plus a JSON schema, an 800-word invoice or form as text, 150 words of JSON out, 100 a month.
- Chat assistant: A 1,200-word system prompt with product facts, 8 turns per conversation, 40-word questions, 150-word answers, 20 conversations a month.
- Agent with tools: Instructions plus 10 tool definitions (2,500 words), 6 model calls per task, 100 words out and a 500-word tool result per step, 40 tasks a month.
| Feature | Claude Opus 5.5 | Claude Sonnet 5.5 | Claude Haiku 5.5 | GPT-6.1 Sol | GPT-5.4 mini | Gemini 3.8 Flash (2026 price) | Gemini 3.1 Flash-Lite |
|---|---|---|---|---|---|---|---|
| Summarise a document | $0.63 | $0.31 | $0.02 | $0.26 | $0.11 | $0.10 | $0.03 |
| Extract fields from a document | $1.10 | $0.55 | $0.03 | $0.57 | $0.23 | $0.22 | $0.08 |
| Chat assistant | $1.16 | $0.58 | $0.03 | $0.45 | $0.20 | $0.42 | $0.15 |
| Agent with tools | $2.15 | $1.07 | $0.06 | $0.83 | $0.35 | $0.51 | $0.18 |
What caching and batch do, on Claude Sonnet 5.5
| Feature | No caching | Caching on | Caching and Batch |
|---|---|---|---|
| Summarise a document | $0.34 | $0.31 | $0.16 |
| Extract fields from a document | $0.75 | $0.55 | $0.27 |
| Chat assistant | $1.47 | $0.58 | not suitable |
| Agent with tools | $3.79 | $1.07 | not suitable |
Prices from the official pages of Anthropic,OpenAI, andGoogle, checked 2026-10-13. Per user per month, model tokens only. Different tokenizers turn the same text into different token counts, so treat cross-provider numbers as approximate.
How it works
Each feature is a run of model calls. A summary or an extraction is one call. A chat is one call per turn, and an agent is one call per step. Every call sends the shared instructions, the input, and everything said so far, so the prompt grows with each turn or step.
With caching on, each call reads the part it already sent from the cache at the cache price, and pays full price only for the new part. Words become tokens at OpenAI's rule of 100 tokens per 75 words, plus 30% on Claude 4.7 and later models.
What it leaves out
Hosting, logging, vector search, Stripe fees, retries after errors, and paid tool calls such as web search ($10 per 1,000 searches on Claude and OpenAI) are not in the number. Neither are long-context surcharges. Caches can also go cold when traffic is low.
To turn model cost into a gross margin per plan, with Stripe and infra included, use theAI SaaS margin calculator.
Questions clients ask
Short answers, no sales language. If yours isn't here, ask an engineer directly.
Ask us anythingHow much does GPT or Claude cost per user per month?
For the four example features on this page, between about one cent and a few dollars per user per month, depending on the model. On Claude Sonnet 5.5 with caching on, a chat assistant used for 20 conversations a month costs about $0.58 per user, and an agent that runs 40 tasks a month costs about $1.07. The same agent without caching costs about $3.79.
Why does a chat cost more per turn as the conversation goes on?
Model APIs do not remember earlier turns. Your app sends the whole conversation again with every message, so turn 8 pays for turns 1 to 7 as input. Prompt caching bills the repeated part at a small share of the input price: 5% on Claude Sonnet 5.5 and Opus 5.5, and 5% to 10% on the OpenAI and Google models listed here.
When should I use the Batch API?
When the user does not wait for the answer: overnight summaries, bulk document extraction, back-office jobs. Anthropic, OpenAI, and Google all list Batch at half the standard price (checked 2026-10-13). OpenAI says each batch "completes within 24 hours (and often more quickly)". Anthropic says most finish "in less than 1 hour".
How accurate is the words-to-tokens conversion?
It is an estimate. OpenAI's rule of thumb is that 100 tokens is about 75 words of English. Code, tables, and other languages use more tokens per word. For real numbers, read the usage object your API returns on each response, or the provider's token counting endpoint.
Adding AI to a product?
We add AI features to products that already have users, with cost caps and usage limits built in. An engineer replies, not a sales team.