What it costs
We measured what it costs to ask an answer engine a buyer question properly, and published the figures. The measurement is cents. The expensive part is everything around it, and almost nobody prices that out loud.
The short answer
We publish our own measured figures below. They are the real amounts billed to our own account on 4 November 2026 for the Anthropic API with web search enabled, not estimates, and they are the only numbers on this page we can vouch for completely.
Our own numbers
One buyer question — “which companies offer X, and what is each strongest at” — put to Claude with web search on, varying only the model and the search allowance. The verdict column is what the measurement was for: whether a particular company was named in the answer.
| Configuration | Cost | Searches | Input tokens | Verdict |
|---|---|---|---|---|
| Top tier, provider defaults | $0.6170 | uncapped | — | not named |
| Top tier, low effort, 5 searches | $0.1684 | 2 | 19,858 | not named |
| Mid tier, low effort, 5 searches | $0.3595 | 5 | 84,356 | not named |
| Small tier, 5 searches | $0.1034 | 4 | 56,802 | not named |
| Small tier, 2 searches | $0.0431 | 2 | 16,908 | not named |
Fourteen times the price, same answer to the same question on the same day. The expensive configuration was not a better measurement; it was measuring an experience no buyer is served.
The counterintuitive row
Look again at the middle row. The mid-tier model is cheaper per token than the top tier and it cost more than twice as much, because it spent its entire search allowance and read four times as many tokens. Every search result is pulled into the input context and billed per token, so reading dominates the bill, not answering. The number of searches is the lever; the model is second.
This has a pleasant consequence. The cheap configuration is also the faithful one, because a person asking an engine a question gets a quick answer rather than a research run. If a vendor’s measurement costs twenty times ours, it is not more accurate — it is measuring a different thing.
Scaling it up
A measurement worth reporting is a set of questions, repeated, across engines. Here is the arithmetic from the per-question figure above, with the caveats that matter.
| Panel | Calls a week | Engine spend |
|---|---|---|
| 17 questions, 1 engine, 1 pass | 17 | $0.59 — measured, not estimated |
| 50 questions, 2 engines, 3 repeats | 300 | roughly $10 |
| 100 questions, 4 engines, 3 repeats | 1,200 | roughly $40 |
Two honest caveats on the lower rows. They extrapolate one engine’s measured rate to four, and three of the four major engines do not publish token rates we have been able to verify, so those figures are arithmetic rather than measurement. And web search is often billed per search on top of tokens, in the region of ten to fourteen dollars per thousand searches, which the rows above include at our own rate but cannot for the engines we have not run.
The expensive part
The question set is the measurement. Getting fifty to two hundred questions that a buyer would really type, in their words, takes a conversation with the people who take those calls. Budget a few hours of their time, once, and revisit it twice a year.
Every claim that gets published has to be traced to something you approve. For a company with assets, registry entries and a decade of releases, this is the longest afternoon in the project and the one that determines whether anything else works.
Correcting a registry entry or a third-party database row is a form, a wait and sometimes a conversation. It is slow, it is unglamorous, and per hour spent it moves more than anything else on this list.
The control group costs nothing to run and something real in discipline: you are declining to improve a fifth of your own questions so that the other four fifths mean something. Every vendor who skips it is saving that cost and charging you for the answer anyway.
How to read all this
Every figure above is the Anthropic API with web search, billed to us on 4 November 2026. Rates change, and a different engine will produce a different table.
These are costs, not what anyone charges. We publish them because the cost of measurement in this category is usually hidden inside a subscription, and knowing it is cents makes the rest of a quote easier to interrogate.
We compared configurations on whether they reached the same named-or-absent answer, not on answer quality. A cheaper model writes a worse paragraph. It did not change whose name was in it.
Questions
Between about two cents and two dollars for the same question, depending entirely on settings. In our own measurements on 4 November 2026, one buyer question put to Claude with web search on cost $0.6170 at the provider's default settings and $0.0431 with a small model and a two-search cap — and both returned the same verdict about which companies were named.
Not for this. A measurement is asking what a buyer would be served, and a buyer is served a quick answer rather than a twenty-search research run. In our runs the cheap configuration returned the same named-or-absent verdict as the expensive one on every question we compared, so the expensive settings were measuring something nobody is shown.
How many web searches the model is allowed, far more than which model it is. Every search result is read into the input context and billed per token, so reading dominates the bill. We measured one mid-tier model costing more than a more expensive one, because it spent its full search allowance and read four times as much.
Cost scales linearly with engines, so start with two and add the others once the question set has settled. The caution is that three of the four major engines do not publish token rates we have been able to verify, so a four-engine budget is an estimate until you have run one call on each and measured it.
Related reading
Nine questions that separate a measurement from a dashboard, with our own answer to each. No vendor is named.
What you get for the discipline of a control group, and the three things it still cannot tell you.
The work the measurement is measuring: what makes a page liftable, and what does nothing.
The definition, and the five things that move whether your name is in the paragraph.
Start here
The free diagnostic: how public sources describe your company and your main product today, where they disagree, and which disagreement to fix first. Every line links to the page we read it from. Nothing is published anywhere without your written approval.