The Free Tier Is Not Dead. Compute Put a Meter on It, and 43% Is the Tell

43% of AI-native products offer a free trial, against 61% of pure software products, in a survey of 200 products run in January 2026. That 18-point gap is the clearest published sign that free access is being rationed. The free tier is not disappearing. It is being converted from an unlimited feature promise into a metered compute budget. The reason is simple. A free user of an AI product now costs real money every time they press the button.

What a free tier always assumed

A free tier is a bet on marginal cost. It works when serving one more non-paying user costs so little that the option value of that user is worth more than the bill.

For twenty years that assumption held. A free Dropbox account cost storage. A free Slack workspace cost a database and some bandwidth. The costs were real but they were pennies, and they fell every year as hardware got cheaper.

AI products broke the assumption in one specific place. The cost is no longer in storing what the user made. It is in the act of using the product at all.

The three costs a classic free tier absorbed

Storage, bandwidth and support. All three scale sub-linearly with users, because most free accounts go dormant within weeks and dormant accounts cost almost nothing.

Inference does not behave that way. A free user who runs one agent task can cost more in a single afternoon than a dormant free account costs in a year. The distribution is also wrong: heavy free users are the ones most likely to be evaluating, automating, or abusing, and all three cost money.

That is why the conversation about free tiers now sounds like a conversation about cloud bills. It is the same conversation, which we have taken apart in more detail in the piece on how inference costs sit inside AI gross margins.

The number that shows the retreat

Most claims about the death of the free tier are anecdotes about one vendor. There is one published figure that measures the direction across a sample.

In January 2026, Kyle Poyar's Growth Unhinged, with ChartMogul and ProductLed, surveyed 200 software products on how they convert free users. 61% of pure SaaS products carried a free trial, against 51% of hybrid SaaS and AI products and 43% of AI-native products.

Read the ordering rather than any single number. Free access thins out in exact proportion to how much of the product is model inference. That is not a coincidence, and it is the only cross-sample evidence in this post.

What the survey actually measured

It measured whether a product offers a free trial, and separately what share of free signups became paying customers within six months. The median free-to-paid rate across all 200 products was 8%.

The spread is the useful part. One in four freemium products converted under 2.5%, and freemium products rarely cleared 15%. Free trials sat higher, with roughly one in four products above 25%.

Two cautions on this source. It is a self-reported survey, not audited billing data, and the sample is skewed toward companies that participate in growth communities. Treat the levels as directional and the ordering as the signal.

Who has already changed, and what replaced it

The direction is easier to see in dated vendor decisions than in survey averages. Five of them, in order, with the replacement stated.

This is a hand-picked list of dated announcements, not a sample. It shows that the pattern exists and is documented. It cannot tell you what share of the market has moved, which is the honest limit of the table.

AWS traded 12 months for $200 and 6 months

On 16 July 2025 AWS announced that new customers receive $100 in credits at sign-up and up to $100 more for using services. The free account plan then ends 6 months after sign-up, or when the credits run out, whichever comes first.

That is a clean swap of a time-based promise for a spend-based one. The old free tier said you may use these services for a year. The new one says you may spend this much, and the clock is shorter.

Accounts created before that date keep the older terms, which is the usual pattern. Grandfathering is what makes these changes hard to see from outside: the population on the old plan shrinks quietly rather than switching all at once.

OpenAI monetised the free tier rather than cutting it

On 9 February 2026 OpenAI began testing ads for US users on the Free and Go tiers, with Plus, Pro, Business, Enterprise and Education staying ad-free. Free users may opt out of ads, and their usage is restricted if they do.

That last clause is the whole argument in one sentence. Free access is now priced, and the price is either your attention or a smaller compute allowance. Both are meters.

The scale explains the choice. OpenAI reported 900 million weekly active users and 50 million paying subscribers in late February 2026. At that ratio, roughly 5.6% of weekly users pay, and the remaining 94% consume inference that has to be funded from somewhere. I would not copy that answer, because ads only work at a user base almost no B2B product will ever reach.

The replacement is a meter, not a paywall

The instinct when free gets expensive is to move features behind a wall. That is the wrong move for an AI product, and the market has mostly worked this out.

A feature paywall does not control cost. A free user with access to one feature can still run it a thousand times. What controls cost is a unit that maps to the underlying bill.

Copilot Free is priced in completions, not features

When GitHub launched Copilot Free on 18 December 2024, the free plan came with 2,000 code completions and 50 chat messages a month. Free users still got a frontier model.

Read what that design does. It gives away quality and rations quantity, which is the opposite of the classic freemium move of giving away quantity and rationing quality. It works because quality is what convinces and quantity is what costs.

This is the single most copyable idea in the whole shift. If you are choosing what to restrict, restrict the number of expensive actions, not the capability. The same logic is why per-seat pricing keeps losing ground, which we covered in the piece on seat compression and what replaced the seat.

The arithmetic you can run this week

Everything above is context. This is the part that decides your free tier.

A free tier survives when the compute it consumes is repaid by the customers it produces. Written as one line, the conversion rate your free tier must clear is:

Break-even conversion = (monthly cost per active free user × months they stay free) ÷ (monthly gross profit per paid user × months they stay paid)

Nothing in that formula is a market statistic. Every input is a number you already own or can pull in an afternoon. That is deliberate, because your inference cost per free user is not a figure anyone else can publish for you.

The three inputs nobody writes down

Most teams cannot answer the first term. They know total model spend and total signups, but not model spend attributable to free users who are actually active.

Split it. Tag inference calls by plan, sum the free-plan spend for a month, and divide by the count of free users who made at least one call. That single number tells you whether you have a marketing cost or a solvency problem.

The table below runs the formula at three cost levels, holding the revenue side fixed. The cost inputs are illustrative placeholders chosen to span a plausible range, not measured figures.

All three rows hold the revenue side constant at $60 monthly gross profit per paid user, 18 months retained, and 6 months spent on the free plan before converting. The cost column is illustrative, not measured. Substitute your own four numbers and the conclusion may reverse.

The third row is the one that matters, because agentic products are exactly where the industry is heading. When the break-even rate lands above what freemium products realistically achieve, a free tier is not a growth engine. It is a marketing budget, and it should be approved and capped like one.

This is also why falling token prices are a weaker rescue than they look. Cheaper units get spent on longer chains of reasoning, which is the pattern taken apart in the post on why cheaper tokens are producing larger bills.

Where this argument is weakest

Three genuine problems with everything above. I would rather state them than have a reader find them.

Fraud killed free tiers before compute did

Heroku announced on 25 August 2022 that free dynos, Postgres and Redis would end that November. The stated cause was not cost of goods. It was that product, engineering and security teams were spending an extraordinary amount of effort managing fraud and abuse of the free plans.

That is three years before anyone was writing about inference margins. It means at least part of the trend I am describing is the older and duller story of free plans attracting bad actors. It is amplified now because those actors can convert a free account into compute they resell.

If you attribute your own free tier decision entirely to model costs, you will design the wrong control. Abuse is fixed with verification and rate limits. Cost is fixed with a meter. They are different problems that produce the same symptom.

Free tiers are also expanding

The single largest free AI offering in developer tooling launched during this period, not before it. Copilot Free arrived in December 2024, after two years of inference cost pressure, and it gave away a frontier model to anyone with a GitHub account.

Cursor is a similar case in miniature. It removed the 7-day Pro trial in January 2026, and its free Hobby plan stayed. What went away was the temporary full-access window, not free access itself.

Both cases point the same way. Well-capitalised vendors treat free access as a distribution weapon and will fund it from other lines, which is a strategic choice about who owns the developer, not an economic one. That distinction is at the centre of the argument about whether to build or buy coding agents.

The third weakness is simpler. The survey behind my headline figure asks about free trials, and a free trial is not a free tier. A product can drop the trial and keep a generous free plan, exactly as Cursor did. My 43% figure is evidence that free access is being rationed. It is not proof that free tiers are being removed, and I have not found a published sample that measures the second thing directly.

Four designs that survive a compute bill

Ranked by how well they control the bill without destroying the reason free exists, which is proof that the product works.

The compute budget is the design I would pick for almost any team under $50 million in revenue. It controls the bill precisely, it never misrepresents the product, and it produces a natural upgrade moment when the allowance runs out.

The reverse trial is second, and it pairs well with the first. Give full access for a fixed window, then drop to the metered plan rather than to nothing. Be careful with the claims made for it: most of the published evidence for its conversion advantage comes from vendors selling the model, so measure it against your own baseline before believing the uplift.

Credits are the packaging layer that carries all of this. In Growth Unhinged's April to May 2026 survey of 230 B2B software and AI companies, 29% already sold AI credits and 33% more intended to introduce them within 6 to 12 months. Hybrid pricing, meaning a subscription with usage on top, rose from 25% to 37% in a year.

One warning on credits. A credit that does not map cleanly to something the user recognises is a tax they cannot forecast, and buyers have started pricing that uncertainty into renewals. Make one credit equal one thing a user can name.

Frequently asked questions

Where to start this week

Start with the number you do not have. Tag every model call by plan, run one month, and divide free-plan inference spend by the count of free users who made at least one call. Most teams discover the figure is either trivially small or alarming, and either answer ends an argument that has been running for months.

Then put that number into the break-even line above and compare the result with your actual free-to-paid rate. If break-even sits above what you achieve, do not debate the free tier in the abstract. Cap it at a monthly allowance, keep the model quality intact, and watch what happens to the conversion rate over one quarter.

If you sell into companies making this same decision, the diagnostic question is short. Ask a vendor what one credit buys. A vendor who answers with a named user action has designed the meter. A vendor who answers with a formula has not.

Related reading