You put an AI assistant on a public page. Every visitor can run it, nobody has to buy anything first, and the bill scales with traffic. The cost of a public AI feature is set by a different model from the one your team built when the feature lived behind a login, and the gap is about to get wider.
The Token Deflation Illusion
Two things are true at once, and only one of them is in most budgets. Gartner expects token costs to fall by 95% by 2030. Gartner also expects inference costs per agentic workflow to rise more than fivefold over the next two years, where an agentic workflow is a task the model plans and executes across several steps rather than answering in one pass. A token is the unit of text a model reads and writes, and inference, the act of running a trained model to produce an output, is billed by the token.
Gartner calls the gap between those two numbers a token-deflation illusion: buyers see the unit price falling and assume the total falls with it. It does not, because the cheaper unit is what makes the more ambitious workflow affordable to attempt. A single question to a chatbot is one model call. An agent that reads the question, retrieves three documents, plans a response, checks its own work and calls a tool is many calls, and each one is longer.
The multipliers are not small. Gartner puts advanced agents at up to 150 times the cost per task of a basic chatbot, and reports that agents consume five to thirty times more tokens than a chatbot for an equivalent task. Falling unit prices do not survive contact with those numbers.
The other half of this story is in our earlier piece on falling frontier model prices. Both halves hold. Per token, the trend is down. Per finished piece of work, the trend is up, and the second number is the one your invoice reflects.
Why a Public Page Is the Worst Place to Learn This
Behind a login, usage has a natural ceiling. Seats are finite, accounts are known, and a customer who runs the feature hard is usually a customer paying you something. The unit economics are legible even when they are bad.
A public web surface removes every one of those bounds. Anyone can run the feature. Nobody has signed anything. And the volume is set by traffic, which is the one variable a web team explicitly does not control: a campaign lands, a link travels further than expected, a competitor sends their whole team to look. Each of those is a good day for marketing and a straight-line multiplier on an inference bill.
Then there is the traffic that is not human. Crawlers, and AI agents running errands for somebody else, now read sites the way people used to, and they do not bounce, get bored, or ask one question and leave. An assistant that answers whatever it is asked will answer those visitors too, at the same cost per task, with no chance of converting. What that traffic looks like is covered in our agentic web readiness framework. The cost consequence is the part that piece did not price.
Four Things to Decide Before the Feature Goes Public
None of these are hard. All of them are much harder to add after launch, once the behaviour is live and someone has to explain a bill.
1. Cache what does not need a model
A large share of what visitors ask a public assistant is the same handful of questions, asked slightly differently. Those answers do not change between visits, so they do not need a model run each time. Resolve the common intents to stored answers and reserve inference for what is genuinely novel. This is the single largest reduction available and it also makes the feature faster, which is the rare control that improves two numbers at once.
2. Gate the expensive path behind stated intent
Not every visitor needs the agentic path, and the agentic path is where the 150x sits. Let the cheap path answer first, and open the expensive one only when the visitor does something that signals real intent: asks a follow-up, supplies a detail about their own situation, or asks for something the cheap path plainly cannot do. A visitor who will not spend one extra click is not worth thirty model calls.
3. Meter per session, not per month
A monthly budget alert tells you about the overrun after you have paid for it. A per-session ceiling stops it while it is happening. Decide how many model calls one visit is allowed, decide what the feature does when it hits that number, and build the second thing at the same time as the first. A graceful handoff to a contact form or a stored answer is a fine outcome. A silent error is not.
4. Keep non-human traffic off the agentic route
Your existing bot rules were written to protect content and rankings, not to protect a metered endpoint. Review them against the new question of what it costs when something automated runs your feature a thousand times. Serve crawlers the readable, cached version of your answers, which is what they actually want, and keep the model calls for visitors who might become customers.
What a Public AI Feature Costs Per Visit
The number that resolves all four decisions is one your marketing team already has: the value of a session on the page where the feature lives. Put the ceiling below it and the feature is a cost of doing business. Put it above and you are paying for attention you were getting anyway. Most teams have never compared the two, because the AI feature was scoped as a product decision and the traffic value lives in a different report.
Key Takeaways
- Token prices and workflow costs are moving in opposite directions. Gartner expects token costs down 95% by 2030 and inference cost per agentic workflow up more than fivefold within two years.
- Advanced agents run up to 150 times the cost per task of a basic chatbot, and consume five to thirty times more tokens for equivalent work.
- A public AI feature is billed by traffic, which a web team does not control, rather than by seats, which it does.
- Caching the common intents is the largest single reduction available, and it makes the feature faster at the same time.
- Set the per-session ceiling against what a session on that page is worth, a number marketing already has and product usually never sees.
Frequently Asked Questions
Does this mean we should keep AI features behind a login?
No. It means the public version needs a cost design the gated version never required. A login bounds usage for you; on a public page you have to build that bound yourself, out of caching, intent gating and a session ceiling. Teams that do this ship public AI features with predictable margins. Teams that skip it either pull the feature after the first surprising month or leave it running and stop looking at the line item.
How do we estimate the cost before we launch?
Take your current sessions on the target page, assume a realistic share will try the feature, and multiply by the cost of the most expensive path a visitor can reach, not the average path. The average is what you hope for; the expensive path is what you are exposed to. Then run the same arithmetic against your best traffic month rather than a typical one, because the expensive month is the one that prompts the question.
Will falling model prices solve this for us?
Not on the timeline that matters. Gartner has the unit price falling through 2030 and the cost per agentic workflow rising more than fivefold by 2028, so the increase arrives first, because capability grows faster than efficiency. Design as though the unit price stays where it is. If it drops, that is margin rather than a rescue.
Sources
- Computerworld, "AI inference is getting cheaper, but your agents are getting more expensive," 2026. Link.
- Gartner, "Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028," 2026.
Next Steps
If your team is scoping an AI feature for a public page and nobody has named the ceiling for a single visit, the launch is a bet on traffic staying flat. Stable Solutions designs the caching, intent gating and metering alongside the feature, so the unit economics are a decision rather than a discovery. Explore our App and Web Development services or contact our team to price a feature you are planning to put in front of everyone.