“It depends” is a true answer and a useless one. Here’s what it actually depends on, and roughly how much each factor matters.
The one-line version
You pay per token, and $1 buys 500,000 of them. A typical answered question costs under 2,000 — so the 300,000 free on signup covers around 150 questions, and most small sites never pay at all.
You pay for tokens, not questions
AI providers bill by token — roughly three-quarters of a word. Every request has two halves: the input (your system prompt, the retrieved passages, and the visitor’s question) and the output (the answer). IrisAI counts both against your balance:
Cost per call = input_tokens + output_tokens
On the hosted service, $1 buys 500,000 tokens at every price point — plans and one-off top-ups work out the same per token. A single answered question typically lands under 2,000 tokens all in.
That matters less than most people expect, though, because the input side dominates. The passages you retrieve on every request usually outweigh the answer several times over — so the biggest lever is how much context you send, not how long the reply is.
The four things that move your bill
Maximum response length
A useful lever, and the easiest one to reach. For a site assistant, 800–1,000 tokens is generous — that’s several paragraphs. Limits above 2,000 are for content generation, not answering questions, and they mostly buy you rambling.
Which model you pick
In BYO mode this is your biggest lever: per-token prices across a provider’s line-up can differ by more than an order of magnitude between the flagship and the economy model. In Proxy mode it matters less than you’d expect, because credits are charged on token counts rather than on which model produced them.
Either way, for extractive question answering — where the facts are already in the context and the model is mostly rephrasing — the cheaper models perform close to the expensive ones. Start economical and only move up if your evaluations show a real gap.
How much context gets retrieved
Every retrieved passage is input tokens on every request. More context isn’t automatically better: past a certain point you’re paying to send the model text it doesn’t need. Well-structured content, chunked at headings, means fewer and more relevant passages.
Caching
Identical repeat questions are served from cache and cost nothing. It’s an exact match on the wording with a one-hour window, so “what are your hours” and “what are your opening hours” are cached separately, and re-indexing or editing your system prompt clears it. Useful on a busy page where the same question arrives repeatedly; not something to budget around.
Indexing is a separate, one-off cost
Building your index means generating an embedding for every chunk. There’s no generated answer and no retrieved context, so an indexing call is far cheaper than a chat call — and you pay for it once rather than per visitor. Editing a post re-embeds only that post.
Do budget for it on a large site, though. Every chunk costs input tokens at the same rate a question does, so indexing several thousand pages is a real one-off spend rather than a rounding error.
The exception is a full re-index of a very large site, which is worth doing deliberately rather than casually.
Estimating before you commit
Take your monthly sessions, assume a low single-digit percentage will use the assistant, and assume two to three questions each. A site with 10,000 monthly visits might see a few hundred conversations — not tens of thousands. Most people overestimate this considerably.
Then watch the real numbers. The Account tab shows requests, tokens and credits consumed, so after a week of live traffic you can replace the estimate with a measurement.
Two ways to pay
In Proxy mode you get 300,000 tokens free on signup and 50,000 more every month, then buy a plan or a top-up if you outgrow that. Everything runs through one connection, with usage tracking and balance alerts included. In BYO mode you bring your own provider key and pay that provider directly — no proxy fees, and you manage the billing relationship yourself.
Proxy suits people who want one bill and a dashboard. BYO suits people who already have provider accounts and want direct control.
Where the money actually goes
| Lever | Impact | What to do |
|---|---|---|
| Retrieved context | Highest — every passage is input tokens on every request | Structure content so fewer passages match better |
| Response length | Moderate, and the easiest to change | 800–1,000 tokens for a site assistant |
| Model choice | Large in BYO, small in Proxy | Start economical; move up only if evals justify it |
| Caching | Free repeats within the hour | Nothing to configure |
Start with free trial credits and measure your own numbers.