TL;DR (July 2026): A screenshot recipe went around this month for reverse-engineering what a "$29/month AI tool" actually pays: find the product on Product Hunt, screenshot its pricing tier, find which model it runs ("powered by Claude", "built on GPT"), open the provider's per-token page, and do the math. For most single-purpose wrappers the raw inference cost of a typical monthly user is cents to low single-digit dollars, which is why $29 to $99 tiers exist. That gap is not automatically a scam - it pays for the product, the prompt engineering, support, and the users who do run up real token bills. But it means two things: as a buyer, heavy use is where a flat wrapper price stops being a deal; as a builder, a flat price is a bet that your average user stays cheap, and the meter underneath is the only thing that tells you when the bet is losing.
The viral version of this was a four-step "here's what it actually costs" thread, and the reason it landed is that the math is genuinely simple - the numbers just are not usually put side by side. A wrapper charges you a flat monthly price. Underneath, it calls a frontier model that bills per token. The difference between those two numbers, minus the wrapper's other costs, is the gross margin. You can estimate it in about five minutes.
The five-minute estimate
- Find the tier. Take the plan you would actually buy - say $29/month.
- Find the model. Most wrappers name it, or you can infer it from latency and quality. Say it is a mid-tier model at roughly $3 input / $15 output per million tokens.
- Estimate your token footprint. A chat-style task is often 1,000 to 5,000 tokens round-trip. A heavy document or coding task can be 20,000 to 100,000. Multiply by how many you run a month.
- Do the arithmetic. 200 tasks a month at 8,000 tokens each (say 5,000 in, 3,000 out) is 1M input + 0.6M output = about $3 + $9 = $12 of raw inference against a $29 price. A lighter user at 40 short tasks costs the wrapper well under a dollar.
- Remember what the margin buys. Prompt engineering, retrieval, the UI, support, failed generations, retries, and the free tier all come out of that gap before it is profit.
Two honest caveats keep this from being a "gotcha." First, wrappers get discounts you do not: batch APIs cut input and output roughly in half, prompt caching cuts repeated-context reads by up to 90%, and volume commitments cut list price further - so their real per-token cost is often below the public number you screenshotted. Second, averages hide the tail: a wrapper's margin math only works because the light users subsidize the heavy ones. The whale who runs 3,000 tasks a month is often unprofitable at $29, which is exactly why so many "unlimited" AI tools quietly added usage caps in 2026.
What this means if you buy AI tools
The flat price is a great deal right up until you become the whale. If a tool is core to your workflow and you use it hard, price the alternative: calling the model yourself through a thin harness. Sometimes the wrapper's prompt engineering is worth the markup; sometimes you are paying $99 for a system prompt you could write. The tell is your own volume - light users should keep the flat plan and stop worrying; heavy users should do the per-token math on their actual usage, not the average. This is the same reasoning behind why Claude Code and Kimi Code moved to task quotas instead of raw token billing - the flat number is easier to sell precisely because it hides the tail.
What this means if you build AI tools
A flat price is a leveraged bet on your usage distribution. It works while your median user is cheap and your whales are rare. It breaks silently: you never get an alert that says "margin inverted," you just watch the provider invoice climb faster than MRR. The three levers that keep a flat wrapper solvent are the boring ones - meter every call so you can see per-user cost, cap or tier the heavy accounts before they eat the month, and decide deliberately whether you pass through, mark up, or eat the variance. We laid out those three rebilling models in pass-through vs markup vs flat; the point here is that you cannot choose intelligently without the meter, because the flat price is only safe if you can see the exact moment it stops being.
The uncomfortable version: if you cannot answer "what did my most expensive 1% of users cost me last month?" in under a minute, you are running the flat-price bet blind. That is the number that decides whether the $29 is a business or a slow leak.
The honest take
Reverse-engineering a wrapper's margin is a useful literacy, not a scandal. Most of the gap is real work, and the discounts you cannot see make the raw cost lower than the screenshot suggests. The actionable part is symmetric: buyers should do the math on their own heavy usage before assuming the flat plan is a bargain, and builders should meter per-user cost so the flat plan never quietly inverts. In both cases the answer lives in the same place - the token meter under the price, not the price on the pricing page.