With a bring-your-own-key plugin there are two costs: the software licence, which is fixed, and the AI usage, which is not. The second one is the one nobody explains, so here is the arithmetic.
What one answer costs
A grounded answer sends the model roughly 1,500 to 2,000 tokens — the retrieved passages from your site, the instructions, and a little conversation history — and gets back 200 to 400 tokens.
On a small, fast model that is a fraction of a cent. On a flagship model it is perhaps ten to twenty times more. Both are small; the multiplier only matters at volume.
A worked example
Say 500 questions a month, which is a busy small-business site:
- On a small model: roughly the price of a coffee for the month.
- On a flagship model: still well under a typical hosted chat subscription.
- With a cache serving repeat questions: about a quarter cheaper again, because the same twenty questions come up constantly.
You can run the numbers for your own traffic with the estimator on our pricing page.
What actually drives the bill up
- A flagship model for simple questions. The most common overspend by far. Opening hours do not need a reasoning model.
- Too much context. Sending six long passages per question when three short ones would answer it triples the input cost.
- Re-indexing everything repeatedly. A well-built plugin only re-processes content that changed; if yours does not, a nightly sync is a recurring bill.
- No cache. Answering the same question fifty times at full price.
- Abuse. A public endpoint with no rate limit is somebody else’s free AI.
How to keep it boring
- Start on a small model and only upgrade if answers are visibly weak.
- Set a hard daily spend cap.
- Turn on rate limits per session and per IP.
- Watch the usage screen for the first fortnight, then stop thinking about it.
For most small-business sites the AI itself costs a few euros a month. The thing to protect against is not the per-answer price — it is an unmetered endpoint with no cap, where a scripted abuser can run up a bill overnight.
Questions
Is there a free option?
Some providers give new accounts starter credit, and embedding models have generous free tiers. You can genuinely evaluate a chatbot for nothing before spending anything.
Why not include AI in the licence?
Because then we would have to guess your usage, add a margin, and cap you. Passing the provider bill straight through is cheaper for you and simpler for us.