When I wrote the book, the question was "what does it cost to serve one more customer?" For most software the honest answer was "almost nothing."
AI changed that. Every summary, draft and agent step sends a meter spinning at a model vendor. So before you set a price, you need to know what that meter will read at the end of the month.
That's what this calculator does. Describe your workload once and it prices it on 20 models at the same time.
How to use the AI token cost calculator
- Pick a preset that looks like your feature (chat reply, RAG answer, document summary or agent run), or type your own token counts.
- Enter requests per month. Count every call to the model. An agent that takes eight steps is eight calls, which is why the agent preset has much bigger numbers.
- Set your cache share. If you resend the same system prompt or reference documents on every call, part of your input can be served from cache at a fraction of the price.
- Tick batch if the work can wait up to a day. OpenAI, Anthropic and Google take 50% off batch jobs.
- Sort the table by monthly cost, cost per 1,000 requests or model name. Add your own model if you have a negotiated rate.
Don't know your token counts? Log a week of real requests. A rough rule is that one token is about three-quarters of an English word, so a 1,500-word document is about 2,000 tokens.
How the maths works
Model vendors charge separately for the tokens you send (input), the tokens the model writes (output) and, on most models, input that's read from cache.
Cost per request = [Input × (1 − cache share) × input price
+ Input × cache share × cache-read price
+ Output × output price] ÷ 1,000,000
Monthly cost = Cost per request × requests per month
Prices are in US dollars per million tokens. Batch, where ticked, halves the result for the three providers that offer it. The tool doesn't add cache-write fees, long-context surcharges or web search fees, so treat the answer as the floor of your bill, not the ceiling.
A worked example
Say you're adding an "ask our help center" assistant to a SaaS product. Each answer sends about 6,000 tokens (instructions plus retrieved articles) and returns 500. You expect 100,000 questions a month, and 40% of the input is a cached system prompt.
On Claude Sonnet 5.5 ($2 in, $10 out, $0.10 cached), one answer costs:
(3,600 × $2 + 2,400 × $0.10 + 500 × $10) ÷ 1,000,000 = $0.0124
That's $12.44 per 1,000 answers and about $1,244 a month. Run the same workload on Claude Haiku 5.5 ($0.10 in, $0.50 out) and it's about $63 a month. Same feature, twenty times cheaper.
That gap is the whole story of AI margins in 2026.
How to read the result
Look at three things.
The spread. If the cheapest model is 20x cheaper than the one you planned to use, test it. On many simple tasks (titles, tags, short rewrites) small models are good enough.
What drives the bill. If most of the cost is input, caching is your first lever. If it's output, ask for shorter or structured answers before you switch models.
Cost per 1,000 requests. This is the number to hold up against your price. If a customer pays $49 a month and makes 600 requests, your inference cost needs to sit well under that, with room for hosting, support and payment fees. Our AI gross margin guide puts the average AI product at about 45% gross margin in 2025 (ICONIQ, via SaaStr), against 75–80% for traditional SaaS.
One more warning. Model prices change. Google's pricing page says Gemini 4 Argon moves from $2 / $10 to $4 / $20 on 1 December 2026. Cost your product at tomorrow's prices, not today's.
What to do next
- Turn this bill into a margin: the AI Gross Margin Calculator adds light, median and heavy users and your plan price.
- See every model's rate side by side in the LLM API pricing comparison.
- Learn how vendors count and charge tokens in token-based pricing explained.
- If you resell AI to customers, design a margin-safe rate card with the Credit Pack Designer.
- Australian business paying in US dollars? Currency swings move your margin, so build in a buffer.
Frequently asked questions
How many tokens is a word?
For English text, one token is roughly three-quarters of a word, so 1,000 words is about 1,300 tokens. Code, numbers and other languages often use more tokens per word. The most accurate way to know is to log the token counts your model vendor returns on each real request for a week and use the averages here.
Why are output tokens more expensive than input tokens?
Writing is harder work for a model than reading. Each output token is generated one at a time, while input is processed in parallel. That's why most vendors charge four to five times more for output. It also means a feature that writes long answers can cost more than one that reads long documents.
What is prompt caching and how much does it save?
Caching lets the vendor reuse input it has already processed, such as a long system prompt or a reference document you send with every request. Cached input is billed at a much lower cache-read price, 5% of the normal input rate on many of the models in this calculator. Some vendors also charge a small fee to write to the cache.
Does this include the cost of AI agents?
Yes, if you count each step. Agents call the model several times per task and resend their growing context on every step, so one agent run can cost many times a chat reply. Use the agent preset, or multiply your requests by the average number of steps per task, to get a realistic monthly figure.
Are these prices in US or Australian dollars?
US dollars, because that's how model vendors bill. If you report in Australian dollars, enter today's exchange rate in the optional box and the calculator will show the A$ cost alongside. Remember that a weaker Australian dollar raises your costs even when your usage stays the same.
Built on The Subscription Playbook
This tool turns frameworks from Robert Coorey's book into something you can use in minutes. Robert is the author of The Subscription Playbook and a Director at BPAG Accountants. More about Robert.