Tools / AI Cost Calculator
AI Token Cost Calculator
Turn a usage log or a workload guess into a monthly bill: edit the rates, add cache and batch discounts, compare models and find the break-even point against a flat subscription. Everything runs in your browser, and your rate table stays on your machine.
Rate table
Prices per 1M tokens. The defaults are examples, not a live price list.
| Model / plan | $ / 1M input | $ / 1M output | $ / 1M cached |
|---|
Example placeholders only — replace them with the rates on your own invoice or pricing page, and re-check them after every plan change. This page never fetches prices, so nothing here can silently go stale in your favour.
CSV, TSV or plain text. Columns: model, input tokens, output tokens, cached tokens (optional). Numbers may use commas or k/M suffixes.
Paste a usage log and press Calculate.
Uses the workload above (requests per day, tokens per request, cache and batch settings) and applies it to every row in the rate table.
Why AI bills are hard to predict
Three things make token spend jumpier than a normal hosting bill: input tokens usually dwarf output tokens and are priced differently, cached input has its own rate, and the same feature can swing by an order of magnitude depending on how much context you attach to each request. A cost calculator is mostly a way to make those three effects visible before the invoice arrives.
How to use it
- Start by replacing the example rates with the ones on your invoice, or with the pricing page of the model you actually call. Rates are stored in your browser and never uploaded.
- If you already have numbers, paste a usage log — one row per model with input and output tokens — and the calculator will total it per model, flagging any model name it cannot match to a rate.
- If you are still designing the feature, use the workload tab: tokens per request, requests per day, cache hit rate. The comparison tab then shows what the same workload costs on every rate row.
Where the money actually goes
In most chat and agent features the biggest lever is not the model choice, it is the context: trimming retrieved documents, capping conversation history and caching a stable system prompt usually beats switching to a cheaper model. Output length is the second lever, because output tokens are typically several times the price of input tokens.
FAQ
Are the default rates accurate?
No — they are examples with round numbers so the calculator is usable the moment it loads. Provider pricing changes, and this page deliberately never fetches anything. Replace them with your own figures and the arithmetic will be correct for your case.
How accurate is the token estimate from text?
It is a rough heuristic: about one token per CJK character and roughly one token per four characters of English. It is good enough to sanity-check an order of magnitude, not to predict an invoice — for that, read the token counts your API returns.
When is a subscription better than paying per token?
When your monthly API cost is above the subscription price. The comparison tab turns that into a break-even figure: how many requests per day your workload must exceed before the flat plan wins — and remember that heavy users often hit usage limits before they hit the break-even point.
Is my usage log uploaded?
No. Parsing, arithmetic and export all happen in the page. Internal model names and request volumes stay on your machine, which is usually the reason people prefer not to paste cost data into a random web form.