Does an AI App Need a Free Tier?

Free tiers were invented for software with no marginal cost. One more user on a note-taking app costs a rounding error. One more user on an AI app costs inference, every time they press the button, whether or not they will ever pay you.

That changes the question. Not "should we have a free tier" as a growth philosophy, but "what is the most we can afford to spend on a stranger, and does a free tier buy more than that is worth". Here is how to get a number before you launch rather than after.

The number you need first

Three inputs, all of which you can estimate badly and still be better off than guessing.

• Cost per free user per month. Your average tokens per active free user times your model price, plus anything else that scales per user. Be pessimistic. Free users are not light users, they are often the heaviest, because they are testing.

• Free to paid conversion rate. If you have no data, assume 2%. For self-serve products that is a normal figure and the common error is assuming 10%.

• Gross margin per paying customer per month. Price minus what it costs you to serve them.

Then:

Maximum affordable cost per free user per month = (gross margin per paying customer x conversion rate) / payback months

Put real numbers through it. Say your paid plan is $20 a month with $14 gross margin, conversion is 2%, and you want the free tier to pay for itself inside six months. That gives (14 x 0.02) / 6, which is about $0.047. Under five cents per free user per month.

That is a tight budget. It is roughly a few hundred thousand tokens on a cheap model and nothing at all on a frontier one. Most free AI tiers that feel generous are losing money on purpose and calling it marketing, which is a legitimate choice as long as it is a choice.

Free tier or free trial

These get used interchangeably and they are different products.

For most AI products the trial is the better instrument, because it caps your exposure per user instead of accumulating it. A fourteen-day trial with a generous allowance usually costs less in total than a permanent tier with a stingy one, and it converts better, because scarcity creates a decision date.

The exception is when the free users are the product: they generate content, data, or network effects that make the paid experience better. If that is genuinely true, a permanent tier is an input cost rather than a marketing cost, and the maths above is the wrong maths. Be honest about whether it is true. Most products say it and few can show it.

Make the limit the thing they would pay for

If you do run a free tier, the limit should sit on the dimension that correlates with value, not on an arbitrary count.

A limit on documents processed per month tells a heavy user they need to upgrade and lets a light user stay forever, which is fine, because the light user costs you almost nothing. A limit on features is worse: it shows people a weaker product and asks them to imagine a better one.

The practical rule is to cap volume, not capability. A free user should see exactly what the product does, just less of it. The related decision of which specific features to gate is covered in deciding if an AI feature should be free or paid; this post is about whether the tier should exist at all and what it may cost you.

Whatever the limit is, it needs a meter the user can see. People do not resent a cap. They resent hitting one they did not know about.

The three signals it is working

Check these at 60 and 90 days. If none of them is true, the free tier is a cost centre with a growth story attached.

1. Free users who convert behave differently from those who do not, in a way you can see early. If converters hit a specific action in week one and non-converters do not, you have a funnel to optimise. If the two populations look identical, the tier is not qualifying anyone, it is just absorbing cost.

2. Cost per free user is at or under your calculated ceiling. Measure it, do not assume it. This is where surprise lives.

3. Free users bring other users. Shares, invites, public output with your name on it. If a free user is a distribution channel, their cost is acquisition spend and should be compared against what you would otherwise pay for a click.

Signal three is the one that justifies an expensive free tier. Signals one and two justify a cheap one.

When to charge from the first minute

Some products should skip this entirely. If your users are businesses, your price is above roughly $50 a month, and the problem you solve is expensive, a free tier mostly attracts people who were never going to buy while handing your inference budget to them. A demo and a short paid pilot work better, and the pilot tells you more, because people who pay something tell you the truth. Running a paid pilot covers how to structure one.

For solo builders in particular, the first paying customer is worth more than a thousand free ones as evidence, and finding that first paying customer is the harder and more informative exercise.

FAQ

What conversion rate should I assume before I have data?

2% for a self-serve product with a free tier. If someone quotes you 10%, ask whether they are counting trials or free tiers, because trials convert far higher and the two are often reported as one number.

Can I add a free tier later?

Much more easily than you can remove one. Removing a free tier is one of the few pricing changes that reliably generates public anger, so launching paid and loosening later is the lower-risk order.

How do I stop abuse of a free tier?

Rate limit per account and require a verified email before any inference runs. Expect some abuse and price it in rather than engineering it away, because the controls that stop a determined abuser also create friction for the honest majority.

Does the falling price of models change this?

It moves the ceiling but not the method. Model prices have been falling steadily, so recalculate your cost per free user each quarter rather than treating the launch figure as permanent.

For the wider pricing picture, see AI monetization strategies.