What your free trial costs in model spend, per customer acquired

A free trial with an LLM behind it spends real money on people who may never pay, and that spend is almost never counted where it belongs. Marketing tracks CAC from ad platforms. Finance tracks model spend as cost of goods sold. Trial compute falls between them: it behaves like acquisition cost, since its purpose is converting a prospect, and it arrives looking like COGS, because it comes on the same invoice as paying customers' usage. The arithmetic that matters divides trial spend by conversions rather than by trials, because the users who churn are paid for by the ones who convert. At an 8% conversion rate every conversion carries 12.5 trials, its own and the 11.5 that went nowhere. There's a second, sharper version of this problem: a trial that lets a user run unlimited generous requests is a line of credit issued to anyone with an email address, and the only defence that works is knowing what each trial account has consumed while it's consuming it. Caps and quotas only bind if something is counting.

• Your CAC number contains ad spend and no compute.
• Trial and paid usage arrive on the same invoice and nothing separates them.
• One trial account consumed more than your median paying customer.
• Nobody can say what a trial costs you before it converts.
• You extended trial limits to improve conversion and never priced the change.

1. Add up last month's model spend from accounts that have never paid.
2. Divide it by the number of trials that converted, not by the number of trials.
3. Compare that figure to your reported CAC and see whether it was in there.
4. Find the single most expensive trial account last month and check what it converted to.

Illustrative example

A modelled 1,000 trials a month. Each trial account consumes 3,000,000 input and 240,000 output tokens over its life, priced on Claude Haiku 4.5 at real rates of $1.00 and $5.00 per million from the price book effective 2026-07-02. A modelled 8% of trials convert. Every volume and the conversion rate are modelled.

Per trial the number is $4.20 and sounds like nothing. Per acquired customer it's $52.50, because each conversion carries 12.5 trials at $4.20, its own plus the 11.5 that went nowhere. The second figure is the one that belongs beside your paid CAC, and it's usually missing from it.

On one key or project, trial spend and paid spend arrive on the same provider invoice with nothing to tell them apart. A separate trial key splits it coarsely, and past that the join is an application-level job that has to happen at capture. That's why this number is so often missing rather than wrong: nobody discarded it, it was never recorded separately. Culpa prices every call and attributes it to the account that made it, so trial accounts aggregate on their own and the distribution across them is visible while it's happening rather than afterwards. Because it also carries what each customer eventually pays, the same ledger produces the figure that closes the loop: trial compute per acquired customer, set against what that customer is worth.

• LLM A/B Test Cost: You Shipped the Winner and Not the Bill
• Free Tier Abuse: What One Automated Signup Really Costs
• LLM Cost Anomaly Detection: The Aggregate Has a Noise Floor
• LLM Cost Forecast Calculator: A Range, Not a Number