The Unknowable Toll

The Unknowable Toll

Before you ask a question, you don’t know how expensive it will be.

Not in dollars — in tokens, in context, in whatever quota is counting down in the background. A simple question might stay simple. Or it might pull in a chain: clarifications, examples, counterexamples, the edge case you didn’t know existed, the nuance that only surfaced because the first answer raised it. You discover the depth of the rabbit hole by falling into it.

This is a structural property of LLM queries, not a user failure. The tool is genuinely opaque about its own cost. The meter doesn’t tell you how far you’ll go before you start.

How It Differs From The One More Query Problem

The One More Query Problem is a tragedy of the commons: millions of individual queries that each seem negligible aggregate into real environmental harm. That’s about collective behavior and the gap between individual and collective reasoning.

The Unknowable Toll is different and earlier: it’s about the individual, before the query, trying to decide whether to ask. You don’t know if this question will cost you 500 tokens or 5,000. You can’t tell in advance. The opacity is pre-query, personal, and structural.

Both concepts live in the territory of AI cost — but OOMQP looks outward at aggregate consequence; The Unknowable Toll looks inward at individual uncertainty before the fact.

The Rational Anxiety

Anxiety about quota is often dismissed as irrational — you’re probably fine, don’t overthink it. But the anxiety is structurally warranted. If you can’t know the cost before you pay it, and the cost comes from a finite budget, caution is just correct reasoning.

Faculty managing institutional API budgets feel this. So do students on free tiers watching their message count. So does anyone building on a per-token plan. The anxiety isn’t about being neurotic. It’s about being in an environment where the meter is hidden and the bill arrives after.

Cost Intuition as a Prompting Skill

Prompting Literacy as Digital Divide describes prompting skill as unevenly distributed along familiar lines of privilege. The Unknowable Toll adds a specific dimension: cost intuition — the learned sense of roughly how expensive a given type of question will be.

You develop cost intuition empirically. You ask things, watch what happens to your quota, adjust. Over time you internalize: “this kind of open-ended reasoning question will spiral; this kind of retrieval question won’t.” The skill is real. But it’s opaque to acquire, because the feedback is delayed and the mechanism is invisible.

This means cost intuition is even harder to teach than other prompting skills. You can explain iteration and persona-setting. You can’t easily explain cost forecasting when the model doesn’t expose it.

The Before/After Asymmetry

The thing you lose is the ability to decide.

After asking, you know exactly how expensive that conversation was. Before asking, you know nothing. This asymmetry means all the decision-making happens under maximum uncertainty, and all the information arrives when the decision is already made.

This is unusual. Most cost structures are visible before commitment: the menu has prices, the invoice is issued before payment, the estimate precedes the project. LLM queries break this norm by design. The product is the rabbit hole itself — its depth is the value, and depth is what you’re trying to know in advance.

Institutional Implications

In educational contexts, this opacity distributes unevenly:

  • Students who ask narrow, focused questions spend less — and may not realize they’re prompting efficiently. The efficiency is invisible.
  • Students who explore, follow tangents, and ask follow-ups spend more — and may be getting more value while burning through quota faster.
  • Faculty allocating per-student budgets can’t predict usage distributions because each student’s curiosity has a different depth profile.

None of this is knowable in advance. Budgeting for AI in institutional contexts requires guessing at a distribution that the tool doesn’t reveal.

Open Questions

  • Can models expose cost estimates pre-query without compromising the answer?
  • Is cost intuition a teachable skill, and if so, how do you teach something learned empirically through invisible feedback?
  • Does opacity serve the provider’s interests in ways that are hard to see?
  • At what point does pre-query anxiety reduce the value of the tool more than the quota limit itself does?

See Also