A price you can put in a budget.
You should be able to approve this once, write it into next year's plan and then stop thinking about it. A meter running behind an AI product makes that impossible, so we do not run one. Here is the reasoning, including the part that costs us.
What a per-token bill measures.
An invoice with tokens on it looks like a price. It is closer to a transcript: a record of what your team did last month, totaled after the fact. That has two consequences and the second one is the expensive one.
The first is that you cannot forecast it. To predict the number you would have to predict your own curiosity, weeks ahead, in a month you have not had yet. How many questions somebody asked, how long the answers ran, whether a competitor made a move that had everyone researching harder for weeks. Finance asks what this will cost and the only truthful answer is a shrug with a chart attached.
The second is that the number moves for reasons unrelated to the value you got. A month where the system reads a lot and finds nothing worth acting on can cost more than the month it hands you the idea that pays for the year. You are billed for effort, in a category where effort is the cheap part.
So what does flat mean in practice?
Three properties, and each one exists to remove a different kind of surprise from the invoice.
One number per brand, and it holds whether the month was heavy or quiet. More research, more drafts and a longer document do not move it.
Nothing to top up, no credit pack to watch drain, no overage line arriving weeks after the behavior that caused it.
Add a brand, add a known amount. You can model next year in a spreadsheet instead of squinting at a usage dashboard.
Somebody carries the variance, and it should be us.
Flat pricing does not make the variance go away. It moves it onto our side of the table. A month where your team reads everything and asks for more is a month our own costs climb, and we absorb the difference. That is a genuine expense and we chose it with open eyes, because the alternative is a customer who quietly uses the product less to keep an invoice down. Rationing the research you bought is a worse outcome than a bad month on our books.
A play card is written before kickoff. You commit to the calls, then you play the game. Settling up afterward is a different sport.
spot-playcard.png1024 × 1024Two ways to pay for a model call.
This is the part that usually stays behind the curtain, and it is the reason flat is survivable rather than reckless. Work that needs an agent, the multi-step passes that read, write and check themselves, runs on a flat-rate seat we already hold. Work that is a plain transform, text in and structured data out, runs metered per call, because that path gives us caps, schemas, caching and batching. Routing each job to the right one is dull engineering with real money in it.
Most of the swing in our own costs gets handled there, long before it could become a line item on yours. We learned that the hard way: for a stretch, a config detail sent work down the metered path that was supposed to run on the flat one, same code, same output, different invoice, no warning anywhere. There is a rule in the codebase now that makes it impossible. Our billing complexity is our problem, and keeping it that way is most of what flat pricing buys you.
Nobody launched anything. The nightly pass ran anyway, the queue was short, two people logged in.
A competitor moved, so your team read everything, asked for more angles and rewrote half of them.
New product, new category, new questions, and everyone in the company suddenly has an opinion.
When the meter disappears, teams start asking more questions.
The habit a metered tool teaches a company is caution. Somebody junior has a hunch about why a product page is not converting, and checking it would cost something nobody has authorized, so the hunch dies quietly. Multiply that by a year. You bought a research system and then trained everyone to use it sparingly, which is the exact opposite of the thing you were paying for. When the cost of one more question is nothing, people ask the strange ones, and the strange ones are usually where the interesting answers live.
The honest limit.
Flat works here because the work per brand has a shape we control. A nightly pass, a known set of sources, a review queue a person actually reads. If what you want is unbounded generation on demand with no ceiling anywhere, this pricing would be the wrong fit, and we would rather say that in the first conversation than quote a number and hope. Adding a brand moves the price by a known amount. Not much else does.
Ask us what your brand would run and you get one number, plus what changes if you add another. More field notes.
Talk to us about your brand