AI Free Tier Economics: When Free Users Cost Money

Free tiers are getting smaller everywhere you look. Rate limits are tightening, flat-rate plans are quietly disappearing, and at least one major lab started testing ads against its free base in February 2026 because the free base had become too expensive to carry for free. In June 2026, Cursor rebuilt its Teams plan around two separate usage pools, one for its own models and one for third-party API calls, which is a polite way of saying the company finally decided to stop absorbing a cost it could not predict.

Everyone is reading this as a pricing story. It is not. It is an accounting story, and the accounting changed underneath an entire generation of founders who learned growth in a world where free users were free.

Here is the durable version of what is happening.

In classic software, a free user cost you almost nothing. Storage was cheap, bandwidth was cheap, and the marginal cost of the ten-thousandth free signup rounded to zero. That single fact is the load-bearing assumption underneath every freemium playbook ever written. Give the product away, let usage compound, monetize the fraction that sticks. The cost of being generous was a rounding error, so generosity was almost always the right call.

In AI products, the marginal cost of a free user is not zero. It is a real invoice that arrives every month, it scales with your success, and it is denominated in tokens you did not charge anyone for. The free tier stopped being a line in the marketing budget and became a line in cost of goods sold. Almost nobody has moved it on their own books yet.

I have built and run products on both sides of that line. The mental adjustment is harder than it sounds, because the old instinct is deeply trained. When usage spikes, the SaaS founder feels joy. The AI founder should feel joy and then immediately open a spreadsheet.

What follows is the framework I use to decide who gets in for free, how much they get, and what I expect to receive in return. I call it the Free Tier Ledger.

Table of Contents

• The Invoice Nobody Priced
• The Free Tier Ledger
• The Three Meters: What a Free User Actually Spends
• The Free Tier Ladder: Five Designs, Ranked
• The Qualification Gate
• What the Conversion Numbers Actually Say
• How the Tightening Actually Happens
• The Contrarian Take: Generosity Is Not a Strategy
• What to Do Monday Morning
• Frequently Asked Questions

The Invoice Nobody Priced

Start with the arithmetic, because the arithmetic is what makes this feel counterintuitive.

Inference got dramatically cheaper. GPT-4-class output fell from roughly $30 per million tokens to under $0.50 per million, a drop of about 95 percent in two years and something close to a thousandfold over three. If you only look at that number, you would conclude that free tiers just got safer. Cheaper input, more room to be generous.

That conclusion is wrong, and the reason it is wrong is the whole point of this essay.

Price per unit fell. Units consumed per user rose faster. A 2023 free user sent a few short prompts and read a few short answers. A 2026 free user triggers an agent that reads a document, plans five steps, calls three tools, retries twice, and writes a long structured output. The same nominal task now costs a multiple of what it cost when the per-token price was sixty times higher. I have watched a single unattended agent run consume more tokens than a heavy human user burned in a week.

The result shows up in margins. Software companies built businesses on gross margins in the 70 to 80 percent range. AI products routinely land in the 40 to 60 percent range once real inference is loaded in. That gap is not a rounding difference. It changes how much runway a given amount of revenue buys, how much a free cohort costs to carry, and how quickly generosity turns into a structural problem.

At the frontier the numbers get vivid. One major lab burned roughly $8 billion on compute in 2025 and is on a path toward something like $14 billion in cumulative losses by the end of 2026. Since late 2022, the large providers priced access below cost on purpose, buying adoption first and planning to fix the economics later, once switching away had become inconvenient. That was a deliberate, well-capitalized strategy.

It is also a strategy you almost certainly cannot copy.

The trap I see founders fall into is inheriting a pricing posture from companies with a fundamentally different balance sheet. A lab with billions in committed capital can price below cost for years as a land-grab. A solo founder with eighteen months of runway who prices below cost is not running a land-grab. They are running out of money slightly more slowly than they are growing.

Free is a purchase order you signed without knowing the quantity. That is the sentence I would tattoo on the inside of the eyelids of every AI founder who is about to launch an unlimited free plan because it worked for the tools they grew up on.

The Free Tier Ledger

The Free Tier Ledger is a way of holding two columns in your head at the same time. On the left is what the free cohort costs you. On the right is what the free cohort returns to you. A free tier is healthy when the right column exceeds the left, and it is a slow leak when it does not. Most founders can describe the right column in vivid detail and have never once measured the left.

The critical structural difference between software and AI sits in how the left column behaves as you grow.

The diagram looks obvious once it is drawn, and yet the implication is one that most teams have not absorbed operationally. In the blue world, a viral spike is unambiguously good news. In the red world, a viral spike is good news and a cash event at the same time, and which one dominates depends entirely on decisions you made months earlier about what a free account is allowed to do.

This is why I stopped thinking about free tiers as a marketing question. Marketing questions get answered with judgment about positioning and funnels. This one gets answered with a number: what is the maximum monthly cost a single free account can impose on me, and am I willing to write that check for every signup I get, including the ones who will never pay and the ones who are actively abusing the plan?

If you cannot answer that question with a specific dollar figure, you do not have a free tier. You have an uncapped liability with a signup form attached.

The healthy version of the ledger has three properties. The left column is bounded, meaning there exists a hard ceiling on what any one free account can cost. The right column is specific, meaning you can name what the free cohort returns beyond a vague hope of future revenue. And the two columns are reviewed on the same cadence, because a free tier that was correctly sized at a thousand users is frequently wrong at a hundred thousand.

The Three Meters: What a Free User Actually Spends

When founders do try to cost their free tier, they almost always count one thing and miss two. The one they count is compute. The two they miss are the ones that tend to hurt more at scale.

I think of every free user as running three meters simultaneously.

The capacity meter is the one that has bitten me hardest. Compute overspend shows up on an invoice a month later, which is annoying but survivable. Capacity contention shows up as a paying customer waiting nine seconds for something that took two seconds last week, and they do not file a ticket about it. They just quietly decide the product is slow and start looking around.

If free and paid requests share one queue, your free tier is not costing you inference. It is costing you retention, and retention is far more expensive than tokens.

The credibility meter is subtler. A free user forms an opinion about your product from the free experience and then tells other people about it. If your free tier is deliberately hobbled to control cost, and you have not been explicit about what it is and is not, the review that gets written is about the hobbled version. You paid inference to manufacture a mediocre impression of your own product. That is the worst possible outcome on both columns of the ledger at once.

The fix is not to make the free tier bigger. It is to make it narrower and honest. A free tier that does one thing completely produces better opinions than a free tier that does everything partially, and it costs less. This connects directly to the argument I made about designing cost into the product before launch rather than after, because the meters you can cap cheaply are the ones you designed for.

The Free Tier Ladder: Five Designs, Ranked

There are only five real free tier designs. They differ in how much they cost and how much information they return, and the ranking between them is not a matter of taste. It follows from the ledger.

Task-Boxed Free is the design I reach for by default, and it is the one I see least often in the wild.

The idea is simple. Instead of giving away a quantity of time or a quantity of credits, you give away one complete outcome. One contract fully reviewed. One dataset fully cleaned. One report fully generated, start to finish, at full quality with nothing held back. Then the free tier ends, not because a timer expired, but because the job you promised to do is done.

Three things happen when you structure free this way.

Your cost becomes knowable in advance. You are not guessing at how much a user might consume in fourteen days. You know exactly what one unit of your core job costs to run, because you have measured it, and you multiply that by one. If you have not measured it, you have a bigger problem than pricing, and the place to start is cost per correct task as your core unit metric.

Your signal gets much stronger. A user who consumed six of fourteen trial days tells you almost nothing. A user who ran one job to completion and came back asking for the second one has told you the most valuable thing a prospect can tell you, which is that the output was worth having again.

And your quality never gets compromised to save money. The most common failure I see is a free tier that routes to a cheaper model, produces a worse answer, and teaches the user that the product does not work. Task-Boxed Free lets you serve the good model to everyone, because you are only serving it once.

The tradeoff is honest and worth stating. Task-Boxed Free produces fewer signups than an open buffet, and if your growth model depends on a very large top of funnel, this design will feel restrictive. That is a real cost. My argument is that it is usually a cost worth paying, because a smaller number of users who finished something beats a larger number who sampled something.

The Qualification Gate

Every free tier should buy you information. That is the thing I want founders to internalize, because it reframes the entire decision.

The instinct most people have is that a free tier buys goodwill. Goodwill is real, but it is unmeasurable and it does not appear on either column of the ledger in a form you can act on. Information does. A free tier that returns a clear, specific signal about who is likely to pay is an investment. A free tier that returns only warm feelings is a donation, and donations should be a deliberate choice rather than an accident of default settings.

So before I ship any free tier, I run it through three gates. If it fails any one of them, it does not ship in that form.

Gate one is the cost ceiling. Can you state, in dollars, the maximum a single free account can cost you in a month, and is that ceiling enforced in code rather than in policy? A limit that lives in your terms of service is not a limit. It is a wish. The enforcement has to be server side, because anything client side will be routed around by exactly the users who cost you the most.

Gate two is signal value. Does the behavior you are giving away actually predict payment? This is where a lot of otherwise sensible free tiers fall apart. If your free plan lets people do the thing that casual users want and gates the thing that serious users need, then heavy free usage tells you nothing about willingness to pay, and you are subsidizing precisely the cohort least likely to convert. Give away the beginning of the serious workflow, not the entirety of the casual one.

Gate three is the upgrade trigger. Is there a specific, predictable moment when the free experience must end and the user has to make a decision? Free tiers without a natural ending do not convert, they accumulate. The user settles into a comfortable equilibrium where the product is useful enough to keep and never urgent enough to buy, and you carry their inference cost indefinitely while they feel perfectly satisfied.

The best upgrade triggers are the ones the user actually wants to hit, because hitting them means the product worked. Running out of credits feels like a punishment. Finishing a job and wanting the next one feels like progress. Same gate, completely different emotional signature, and the second one converts dramatically better.

What the Conversion Numbers Actually Say

It is worth grounding all of this in what conversion actually looks like, because founders routinely plan against a number that almost nobody achieves.

Across products generally, the median free-to-paid conversion rate sits around 8 percent, but that median is misleading because the distribution is close to bimodal. Roughly a quarter of freemium products convert below 2.5 percent. Another 29 percent land between 2.5 and 7.5 percent. And about a quarter reach 10 to 15 percent. Very few products actually sit at the median. You are probably going to end up in the bad cluster or the good cluster, not the middle.

Narrow to B2B software and the picture tightens. Median freemium conversion runs about 2 to 5 percent. One analysis of traditional freemium found an average of 3.7 percent, with variation by sector from around 2.6 percent at the low end to 5.8 percent at the high end. For self-serve freemium, 3 to 5 percent is considered good and 8 to 12 percent is considered excellent.

Now compare the alternatives. Opt-in free trials, where no card is required, convert at roughly 18 percent. Opt-out trials, where a card is required up front, convert at nearly 49 percent.

The honest complication, and I want to state it because it cuts against my own argument, is that freemium generates about twice the signups per thousand visitors that free trials do. Roughly 90 versus 45. Once you account for that difference in top-of-funnel volume, freemium products can produce more paying customers per thousand visitors than trial products, despite the much lower conversion percentage.

That was a decisive argument in classic software, and it is why freemium won the last decade. When the extra 45 signups cost you nothing, of course you take them.

In AI, you have to finish the sentence. Those extra signups are not free. Each one runs three meters. So the correct comparison is not conversion rate versus conversion rate, or even paying customers per thousand visitors. It is paying customers per thousand visitors, minus the cost of serving everyone who did not convert.

Run that arithmetic honestly and the ranking often flips. A model that converts at 18 percent on 45 signups can beat a model that converts at 4 percent on 90 signups once you subtract the cost of carrying 86 non-paying accounts on the second one indefinitely. The break-even depends entirely on your cost per free user, which is exactly the number the ledger forces you to compute.

One more finding worth carrying: hybrid pricing, meaning a subscription floor plus usage on top, has been associated with about 21 percent higher median growth than either pure subscription or pure usage-based pricing. The structural reason is that a floor covers your fixed serving cost while usage captures the upside from heavy users, so neither cohort is mispriced. I unpack the full set of options in the deeper piece on revenue models for AI products, and the related case against seat-based pricing in why per-seat pricing is dead for AI.

How the Tightening Actually Happens

The pattern by which companies discover all of this is remarkably consistent, and watching it play out in public is instructive because you can skip several of the steps.

Stage one is generous launch. The product ships with a free tier modeled on software norms, because that is what everyone knows and because early usage is small enough that the cost feels like a marketing expense. This stage feels great.

Stage two is the quiet realization. Usage grows, the invoice grows faster, and someone finally builds the dashboard that shows cost per cohort. There is often a specific meeting where a number gets said out loud for the first time.

Stage three is the correction, and this is where companies get hurt. The correction is usually right on the economics and badly handled on the communication. Cursor’s move to usage-based billing in mid-2025 is the canonical example. The underlying change was defensible. The rollout was not, users were surprised by charges they had not been warned about, and within a few weeks the company posted an apology, clarified the rules, and offered refunds for the affected window.

Stage four is the mature structure. By June 2026, Cursor’s Teams plan had been rebuilt around two separate usage pools, one for first-party models and one for third-party API calls, along with a new premium seat tier and better admin controls, with the company estimating the changes would lower costs for around 90 percent of teams. That is what a considered answer looks like: separate the costs that behave differently, give buyers visibility, and stop cross-subsidizing.

You can see variations of the same arc across the sector. Free plans narrowing to a small daily allowance of the expensive operation while the cheap operation stays unlimited. Flat-rate plans being retired. A major lab testing advertising against its free tier in early 2026 rather than continuing to absorb the cost of an enormous non-paying base.

The lesson I take from the arc is not that you should be stingy. It is that the correction is coming for everyone who launches generous, and a correction you plan is a repricing while a correction you are forced into is a betrayal. Users forgive a limit that was always there. They do not forgive a limit that arrives after they built a workflow on its absence.

So set the ceiling on day one, even if it is generous. It is far easier to raise a limit later, which users experience as a gift, than to lower one, which they experience as a bait and switch. This is the same logic that governs how long a feature stays defensible before the model layer absorbs it. Decisions that look like they can be deferred are usually decisions that get much more expensive to make later.

The Contrarian Take: Generosity Is Not a Strategy

Here is where I part ways with most of the growth advice I read.

The standard position is that in a crowded market you should be more generous, not less, because attention is scarce and a great free experience is how you win it. Give more away. Let people fall in love with the product. Monetize later.

I think that advice is now backwards, and the reason is that it optimizes for a bottleneck that no longer binds.

When building software was expensive and rare, a free tier was a differentiator. Few competitors could afford one, so generosity bought distinctiveness. That has completely inverted. When anyone can ship a competent AI product in a weekend, everyone has a free tier, and being generous no longer distinguishes you from anything. It just means you have joined a group of companies all subsidizing the same undifferentiated sampling behavior. I wrote about the upstream version of this in what to build when building is free, and free tier design is where that shift shows up on the income statement.

The second reason is more uncomfortable. A generous free tier does not select for your best future customers. It selects for the most price-sensitive segment of your market, because those are the people for whom free is the point rather than a starting position. You then spend real inference dollars deepening a relationship with the cohort least likely to ever pay you, while the buyer who would have happily paid on day one gets the same rationed experience as everyone else.

I have made this mistake. The most expensive free users I have ever carried were sophisticated, heavy, delighted, and completely uninterested in ever paying, and they were expensive precisely because they were sophisticated. They knew how to extract the most value per request. A naive user costs you very little. An expert free user can cost more than a paying customer generates.

The version of generosity I now believe in is narrow and complete rather than broad and rationed. Give one real outcome away entirely, at full quality, with nothing held back. That is genuinely generous, it produces a real opinion of your product rather than a compromised one, and it costs a knowable amount.

The thing you should be generous with is not compute. It is clarity. Tell people exactly what they get, exactly what it costs, and exactly when it ends. Nobody has ever churned because a limit was clearly stated. They churn because it was not.

What to Do Monday Morning

Concrete steps, in order, that you can finish this week.

Compute your cost per free user. Pull last month’s inference spend, split it by paying and non-paying accounts, and divide the non-paying portion by the number of active free accounts. Most founders have never seen this number. Then find your 95th percentile free account, not the average, because the average is a comforting lie and the tail is what actually breaks your budget.

Find your ceiling and enforce it in code. Decide the maximum monthly cost you are willing to carry per free account, then implement a hard server-side limit that stops execution when it is reached. Not a warning email. A stop. If you cannot enforce it today, that is your highest-priority engineering task this week, ahead of whatever feature is currently in progress.

Separate the queues. Make sure paid requests never wait behind free ones. If you have one queue and one rate limit pool, your free users are currently degrading the experience of the people paying you, and you probably cannot see it in any dashboard you have.

Run the three gates on your current free tier. Cost ceiling, signal value, upgrade trigger. Be honest. Most existing free tiers fail at least one, and the most common failure is gate three, no defined ending. If yours fails, you now know what to fix and in what order.

Redesign toward one complete job. Ask what single outcome, delivered fully, would convince a serious prospect that your product works. Price that one job. Make it the free tier. Delete the rest of the free plan. This is the single change with the best ratio of effort to margin recovery that I know of.

Instrument completion, not usage. Stop optimizing for signups and sessions. Track what fraction of free accounts finish one real job, and what fraction of finishers convert. Those two numbers tell you more about the health of your business than anything in your analytics dashboard right now.

Write the limit down publicly. Put the ceiling in your pricing page in plain language, before anyone hits it. This costs nothing and buys you the right to enforce the limit without it feeling like a betrayal later.

Set a quarterly review. A free tier correctly sized at a thousand users is frequently wrong at fifty thousand. Put a recurring calendar item on the day you ship it. The companies that got hurt were not the ones with expensive free tiers. They were the ones who did not look at them for a year.

If you are earlier than this and still deciding whether the product is worth building at all, the sequencing matters. Validate demand before you engineer the meter, using the approach in validating a startup idea in 48 hours, and get the underlying margin structure right using the AI gross margins playbook. Free tier design is a question you earn the right to answer, and it sits inside the broader sequence I lay out in the AI-native founder playbook.

Frequently Asked Questions

Should an AI product have a free tier at all?

Yes, in most cases, but not an open-ended one. The purpose of a free tier in an AI product is to let a serious prospect prove to themselves that the output is worth paying for, which usually requires completing one real job rather than sampling many partial ones. If your product has high contract value and a complex deployment, an assisted pilot with a human in the loop will teach you more and cost you less in aggregate than an open self-serve free plan.

How much should a free user be allowed to cost?

Set the ceiling as a fraction of the gross profit on one converted customer, then divide by your realistic conversion rate. If a paying customer generates $40 of monthly gross profit and you convert 4 percent of free accounts, each free account can justify roughly $1.60 of monthly cost before the cohort stops paying for itself. Use your own numbers, and use the 95th percentile of free usage rather than the average, because the tail is what actually determines your exposure.

Why is freemium harder for AI products than for classic software?

In classic software the marginal cost of serving one more free user rounded to zero, so growth in free users was free. In AI products every free interaction consumes inference, so the cost of the free cohort rises in a straight line with adoption. That moves the free tier out of the marketing budget and into cost of goods sold, and it is a large part of why AI product gross margins commonly land in the 40 to 60 percent range against 70 to 80 percent for traditional software.

What is a good free-to-paid conversion rate for an AI product?

For self-serve freemium, 3 to 5 percent is good and 8 to 12 percent is excellent. Median B2B freemium conversion sits around 2 to 5 percent. Opt-in free trials convert at roughly 18 percent and opt-out trials requiring a card convert at nearly 49 percent, so if your unit economics are tight, moving from open freemium toward a trial structure is usually the fastest available improvement.

Does the falling cost of inference solve this problem over time?

Not on its own. Per-token prices fell roughly 95 percent in two years, but consumption per task rose faster as products moved from single prompts to multi-step agent runs with tool calls and retries. Unit costs falling while unit counts rise faster means total cost still climbs. Plan against tokens consumed per completed job rather than against the headline price per million tokens.

Should I require a credit card for the free tier?

It depends on which cost you would rather pay. Requiring a card cuts signups sharply but raises conversion to close to 49 percent and filters out most abuse, which matters enormously when each signup carries real inference cost. If your product benefits from very large top-of-funnel volume, no card is defensible, but only if you have a hard server-side spending ceiling per account.

How do I tighten an existing free tier without angering users?

Grandfather existing accounts on the old terms for a defined window, announce the change before it takes effect rather than after, explain the reason in terms of serving cost rather than value capture, and make the new limits visible inside the product before anyone hits them. The failures that generated public backlash were almost always failures of notice, not failures of pricing. A limit users knew about is a rule. A limit that appears on their invoice is a betrayal.

What is Task-Boxed Free and why prefer it over a time-limited trial?

Task-Boxed Free gives the user enough to complete one real job end to end at full quality, then stops, rather than giving unrestricted access for a fixed number of days. It is preferable because your cost is knowable in advance, since you are funding exactly one unit of work, and because completion is a far stronger predictor of payment than elapsed time. Time-boxed trials frequently expire before the user has finished anything meaningful, which leaves you having paid for inference and learned nothing.