Field note · 19 September 2026
A free AI tool on your site, without an unbounded bill.
At the prompt sizes a marketing tool uses, one run on a small instruct model costs about $0.0001, and a visitor who uses every generator we offer to its limit costs under a cent. What keeps spend there is the order of the checks in front of the model and, above all, a prepaid balance with a spending cap on the key the tools use.
30 min · Fixed quote before we build · Projects from $2,500
What a run actually costs
Our generators send roughly 600 tokens in and cap output between 700 and 1,200 tokens depending on the tool. On a small instruct model that is about $0.0001 per run. We allow ten runs per tool, across seven tools, so a visitor who uses every tool to the limit costs well under a cent.
| Scenario | Runs | Approximate cost |
|---|---|---|
| One visitor, one tool, one run | 1 | about $0.0001 |
| One visitor exhausting one tool | 10 | about $0.001 |
| One visitor exhausting all seven tools | 70 | under $0.01 |
| Reaching $100 of spend | about 900,000 | roughly 13,000 fully exhausted visitors |
Measured on a small instruct model at our prompt sizes, September 2026. A frontier model is roughly fifty times more per run and changes these numbers, not the shape of them.
No new site sees thirteen thousand visitors run every tool to its limit. Ordinary use is not a budget risk. Automated traffic is, and that is where the controls belong.
Put the free checks first
A check that runs before the model call costs nothing when it rejects a request, so the order of the checks matters more than any single limit. Ours run cheapest first, and the paid call comes last.
- Is this a known generator, with the required fields present?
- Is the source sending requests faster than a person could?
- Did the request pass human verification?
- If an email was supplied, is it a real address rather than a disposable one?
- Is the visitor still within the free allowance for this tool?
- Only then is the model called.
What each layer does
Each layer has one job. Together they keep usage fair for visitors and keep automated traffic away from the model.
| Layer | What it does | Why it is there |
|---|---|---|
| Free allowance per tool | Ten runs per tool, shown to the visitor | Keeps the tools open to everyone while giving each visitor a fair ceiling |
| Burst limit | Slows any single source sending requests faster than a person could | Absorbs sudden spikes before they reach the model |
| Human verification | Confirms that each request comes from a real visitor | Stops scripted traffic, the one realistic source of a large bill |
| Email after the first run | Asks for a work email once the tool has shown what it does | Turns a free run into a lead, and refuses disposable addresses |
| Prepaid balance with a capped key | The provider stops serving once the cap is reached | Fixes the maximum possible spend in advance |
The first four layers keep usage fair and automated traffic out. The fifth guarantees the budget: the provider stops serving at the cap, so the most you can spend is a number you chose in advance.
The hard limit: prepaid credit
Put API access on prepaid credit rather than an invoice, and set a spending limit on the individual key your site uses. The worst case is then fixed in advance, by you: the key stops, the tool shows a clear message, and the bill cannot exceed the cap. Every other control in this article exists so that you never get near it.
- Prepaid credit, never an open invoice, for anything a stranger can trigger.
- A spend limit on the specific key the public tool uses, separate from any key you use internally.
- Output token caps per tool, sized to the answer you actually want, because output is the expensive half.
- A short system prompt. Ours is about 230 tokens and gets sent on every single request.
- An alert at a threshold you would want to know about, well below the cap.
We publish the free allowance, because visitors see it anyway. The other thresholds stay private, and we would advise the same for any public tool: they help nobody except someone trying to work around them.
If you want a tool like this built with these controls in place, that is the work on our AI automation agency page.
Questions, answered.
Do I really need human verification on a free tool?
Yes. It is the layer that stops a script from running your generator in a loop, which is the one realistic way a marketing tool produces a large bill.
Can I just rate limit by IP address?
Use it as one layer among several. A per-address limit absorbs bursts well. For a quota that holds across all your servers, keep the counter in shared storage, and for a hard ceiling on spend, rely on the cap at the provider.
Is a cheaper model the answer?
It helps, and the spread between a small instruct model and a frontier one is roughly fifty times per run at these prompt sizes. But choose on output quality and on whether the model answers directly rather than reasoning first, then take the cheaper option among the ones that work.
What should happen when someone hits the limit?
Say plainly that the free allowance is used and offer a next step, such as a conversation with your team. A clear message costs nothing to serve and leaves the visitor with somewhere to go.
Should the tool ask for an email?
After the first run, yes, if the tool exists to generate leads. Asking before anyone has seen it work wastes the demonstration. Asking once it has produced something useful is a fair trade, and it doubles as a cost control.
Related services
Tell us what's slowing you down. We'll show you what to automate first.
Free · 30 min · Projects from $2,500
contact@aistrikeforce.com