The same cut on every model

Technical information first: there is exactly one rule behind every number on the configurator, and it fits in a sentence. Take what one grounded answer costs us on the model you chose, measured on our own account, add the few cents it costs to have a machine read that answer back, and divide by one fixed share; that's the per-question rate for one pass on that rung. The $2 line for drafting the questions and diagnosing the losses is the same rule applied to the work that doesn't scale with the engines. Add the lines up, round to the next $5, never below $25, and that is the price. There's no second sheet, no rate card for large accounts, and no number on the page that came from anywhere else.

We've already written about what a grounded answer costs and about showing the price before you pay. This post is about the part in between, the share, because it's the decision that makes the other two honest. One share, held the same across five engines and twenty-odd model versions, means the page cannot steer you. A line is dearer because the model behind it reads more of the web per answer, and for no other reason.

Why one share and not a menu

Most software pricing steers. The plan the vendor wants you on is the one that happens to be a bargain, the configuration that actually suits you is oddly dear, and the arithmetic that produced both stays on the vendor's side of the table. We'd rather not sell that way, so we removed the only tool it needs: a margin that varies by line. GPT-6 Astra costs more on our page than GPT-4o mini because it issues five searches an answer and 4o mini issues one; Claude Sonnet 5 costs more than Claude Opus 5 because Sonnet reads over a hundred thousand tokens of retrieved context per answer and Opus reads less; Perplexity's Sonar costs under a cent an answer because that is what Perplexity charges for one. Our share on each of those is identical, so when you ask us which models to switch on, the honest answer costs us nothing to give: the ones your buyers are actually served, which is a question about your market and not about our margin. Which model judges you is the longer answer.

An opinion we'd defend in front of anyone who prices software for a living: a vendor whose margin varies by line has a reason to recommend one line over another, and a buyer can never fully trust the recommendation. Flattening the share is how we made our own advice worth taking.

What the rule makes of our own ladder

The rule prices our fixed reports too, so you can check it. Switch on GPT-5.6 Sol, Claude Sonnet 5 and Gemini 3.6 Flash at 20 questions and three passes, and the arithmetic comes to $177.85, which rounds up to $180: the same $180 the fixed ladder sells that report at, because it is the same report. The rail shows the pieces: $91 for the Sol line, $78 for Sonnet 5, $7 for Gemini, $2 for the questions and the diagnosis.

Now the admission, since you'd find it anyway. Add Perplexity's Sonar and Grok 4.3 to that shape, all five engines at the rungs every paid tier answers on, and the arithmetic says $200. The page sells it at $250, because that is what Full Coverage has cost on the ladder since before the configurator existed, and one spec gets one price whichever door you came in through. The $250 is a pin, and a pin is a floor: the formula still wins whenever it lands higher, so a pinned report can never slip under its margin, and it is the single line on the page that isn't the formula. We'd rather say that here than have you notice it on a spreadsheet, and if the ladder price moves, the pin moves with it.

What one share costs us

Two things, and both are on the page if you look. First, a cheap model is a cheap report. A 20-question report on Gemini 3.6 Flash alone clears its margin well under $25 and sells at the $25 floor, which is the one place positioning overrules the rule: we'd rather not sell a report for the price of lunch, and the floor says so in the same file the server reads. Second, we price from the top of the cost band, not the middle. On OpenAI the model decides how many searches to run, four samples of the same job on Sol cost $0.208, $0.253, $0.311 and $0.371 an answer, and the rate is built from the last of those. On a quiet run we keep a little more than the share; on a busy one, a little less; either way you paid the number that was on the page before the model made up its mind.

The share itself is set by hand, by us, and we've moved it. It was one number on the morning of the 8th of September and a different one by that evening, and every price on the site dropped with it; it will move again if a vendor's prices do. What won't move is that it's one number, applied to every line, in one file that both the page and the checkout read. A change to it is a change to every price on the site at once, in public, and that is the constraint we wanted on ourselves.

Where the proof lands

Open the configurator, switch on the engines your buyers use, pick a rung for each, and watch the rail. Move Claude from Sonnet 5 to Opus 5 and the Claude line roughly halves, because the cost did. Add Grok 4.6 and the line says what a model that reads a hundred thousand tokens of search results per answer costs, with our share on top and nothing else. The Stripe receipt names the same models the rail did, at the same total, because the server priced your choices from the same catalogue when you paid.

If you'd rather we chose the engines for you, book a call and we'll pick the ones your buyers are served on, which, at one share, is the only reason we have for picking anything. The report lands with every model named on every stored answer, and the next one, on the same models, costs the same again.