Are Your Free Users Costing You Money? Ours Cost $8k/Month

We launched a free tier to grow signups. It worked — we went from 200 to 2,000 users in six months. Then our OpenAI bill tripled. We assumed it was growth. It was. But it wasn't the growth we wanted.

The math we weren't doing

Free users had the same AI access as paid users. Why wouldn't they? We wanted them to experience the product. The problem: they had no reason to hold back. Power users on the free tier were generating 5,000+ requests a month. Our paying customers? Maybe 500. We were spending more to serve people who paid nothing than people who paid $99/month.

We'd built the free tier as a funnel. Try the product, see the value, upgrade. In theory it made sense. In practice, a chunk of free users were never going to convert — they were students, hobbyists, or competitors kicking the tires. And even the ones who might convert had no incentive to. Why pay when you get everything for free?

How we found out

We couldn't see it in the OpenAI dashboard. We had to connect our usage data to our customer list. Once we did, the numbers were brutal: free users accounted for 58% of our AI spend. Our best-paid customers were subsidising everyone else. The "funnel" was a leak.

The breakdown by feature was just as bad. Our document summarisation tool — the most token-heavy feature — was used disproportionately by free users. They'd upload long PDFs, run summaries, export, repeat. Some were clearly using it for work without paying. We had no idea until we could see cost by customer and by feature.

What we did about it

We didn't kill the free tier. We capped it: 50 AI requests per month. Enough to try the product. Not enough to run a side business on it. We also moved our document summarisation feature behind the Pro plan — that was the biggest cost driver.

We were nervous. Would signups drop? Would people leave? Signups dipped slightly. But conversion from free to paid actually went up. Turns out people who hit a limit are more likely to upgrade than people who get everything for free. The limit created a decision point. "I need more — is it worth paying?" For many, the answer was yes.

Six months later, our free-tier AI spend is down 80%. Our paid conversion rate is up. The free tier is finally doing what we intended: letting people try the product, not letting them run it for free indefinitely.

If you're wondering where your AI spend is going — and whether your free tier is part of the problem — you need the breakdown. We built PerUnit to give you cost by customer, feature, and pricing tier. So you can see the numbers before you make changes. No data engineering required. If you're also trying to reduce costs, our post on how we cut our OpenAI bill by 40% has the tactics that worked for us.