AI tool shutdown rate by pricing model: paid tools close 1.7× as often as freemium

The assumption is so common it rarely gets stated: free AI tools are the fragile ones, and a paid AI tool has revenue and therefore a future. Measure the AI tool shutdown rate across our catalogue and the ranking comes out backwards. As of 13 September 2026, paid tools have shut down at 6.2% and freemium tools at 3.6% — the paid shutdown rate is 1.7 times the freemium one. That is a real gap in a census of 2,760 products, and it is worth taking seriously. It is also smaller than almost anything else that predicts whether an AI tool survives, which is the part most people will skip. This piece gives the numbers, dated, and then tries honestly to explain them — including the explanations the data cannot rule out.

The numbers, dated

Census taken 13 September 2026 across every published listing in the directory. "Shut down" means the product is no longer operating and we have recorded it in the AI Graveyard. "Acquired" is a separate outcome, counted separately, because a bought product is not a dead one and roughly half of ours are still running under the new owner.

Two of those rows carry a warning we want to state before anyone lifts them. Free and free-trial tools each clear our minimum sample size on the denominator, but the numerators are six and ten endings respectively. A shutdown rate built on six events moves by a full percentage point every time one more product closes, so we report those rows and decline to rank them. The finding in this piece is the paid-versus-freemium gap, which rests on 101 and 29 endings — enough to be a pattern rather than a coincidence.

Why paid tools close more often: four explanations, adjudicated

A number like this invites a tidy story. Here are the four we considered, in order of how much the data supports them.

1. The pricing model we record is the last one, not the first. This is the explanation we take most seriously, and it is a limitation of the census, not a fact about the market. We record the pricing model a tool had when we last checked it. A product that launched freemium, struggled, and switched to paid-only in its final months as a monetisation push gets counted as a paid shutdown. If that pattern is common — and anecdotally it is — then part of the paid rate is really the freemium rate wearing a different label. We cannot measure how much, because we do not keep pricing history. Treat the 1.7× as an upper bound on the true gap.

2. Freemium is where the platform features live. A large share of freemium listings are AI features inside products that already had a business — the meeting summariser inside a video-call tool, the writing assistant inside an office suite. Those do not shut down in the way a standalone startup does; they get folded, renamed, or quietly deprecated inside a product that carries on, and our lifecycle field does not always catch that as an ending. Standalone paid products, by contrast, either make money or stop existing, which makes their endings legible. Some of the gap is measurement: paid shutdowns are simply easier to see.

3. Paid tools carry the cost structure. Products that charge from day one skew toward inference-heavy categories — generation, agents, video — where the bill for running the model arrives whether or not customers do. A freemium tool can let its free tier idle cheaply; a paid tool that is not selling is paying for compute it cannot recover. This is plausible and consistent with what we see in the category failure rates, where the riskiest categories are the compute-hungry ones. It is not something this census can confirm on its own.

4. Freemium tools have users, so they get bought instead of buried. This was our first hypothesis and the data does not support it. If a free tier were converting failing products into acquisition targets, the freemium acquisition rate would be visibly higher than the paid one. It is not: 5.7% against 5.9%. Acquisition is flat across every pricing model in the table, within a point. Whatever makes a buyer pick up an AI tool, it is not the presence of a free tier.

The honest summary is that explanations 1 and 2 are about how we count, and 3 is about the market. Somewhere between a measurement artefact and a real cost-structure effect, paid tools close more often. We would not build a strategy on the precise 1.7×. We would stop assuming the number runs the other way.

What the acquisition column actually says

It says buyers do not care about the pricing model. Every bucket sits between 4.0% and 5.9% acquired, and the two large ones are three-tenths of a point apart. Of the 157 acquired tools in the catalogue, 87 are still operating under their acquirer — a survival-through-acquisition rate of a little over half, which we track in more detail in what happens to acquired AI tools.

That flatness is the more useful finding for anyone building a product. Acquirers appear to be buying capability, team, or distribution — things that do not sort by whether the pricing page has a free row.

The comparison that puts this in proportion

Pricing model separates the two big buckets by a factor of 1.7. Category separates tools by a factor of about seven: on the same census date, category shutdown rates run from 2.1% in engineering and simulation to 15.6% in automotive across the 38 categories large enough to rank. What a tool does predicts whether it survives roughly four times more strongly than how it charges.

So if you are using pricing model as a survival signal when evaluating a tool, you are looking at the weaker variable. Look at the category first. Then at how long the company has been operating, whether it has named customers, and whether it is running on its own model or reselling someone else's — the checks in our guide to telling whether an AI tool is legit. Pricing model is real, but it is the last thing on that list, not the first.

How we count

The census is the directory's own catalogue: every published listing, with its lifecycle status and recorded pricing model, read on 13 September 2026. Lifecycle is one of active, shut down, or acquired; shutdowns are confirmed against the product's own domain and announcements before a tool enters the Graveyard, and acquisitions are recorded on announcement with a separate flag for whether the product kept operating. Pricing model is the model recorded at our most recent check, which is the limitation discussed above.

This is a census of listed products, not a random sample of every AI company. It over-represents tools that were good enough, or visible enough, to be listed, and it under-represents the products that never reached anyone. That makes it a better measure of what happens to AI tools people actually use than of the AI startup population as a whole — the distinction that our piece on why every source quotes a different failure rate is about. Read the rates here as "of the AI tools worth listing, how many closed," and nothing broader.

We will re-run this census and update the table; the figures above are frozen to their date and will not be edited in place.

Frequently asked questions

What is the shutdown rate for AI tools in 2026? Across the 2,760 published tools in our catalogue as of 13 September 2026, 146 have shut down — a shutdown rate of 5.3%. A further 157 (5.7%) have been acquired, and 87 of those are still operating under their new owner. The rate varies by category far more than by pricing model: from 2.1% to 15.6% across categories, against 3.6% to 6.2% across pricing models.

Do free AI tools shut down more often than paid ones? Not in our census. Paid tools have shut down at 6.2% and freemium tools at 3.6% as of 13 September 2026, so paid tools close about 1.7 times as often. Part of that gap is likely measurement — we record a tool's last pricing model, so a freemium product that switched to paid before closing counts as a paid shutdown — but the direction is the opposite of the common assumption.

Why would a paid AI tool be more likely to shut down than a free one? The most defensible explanations are about counting rather than the market: we record the final pricing model, and many freemium listings are features inside larger products that get folded rather than formally shut down. The market explanation with some support is cost structure — paid tools cluster in inference-heavy categories where compute costs arrive regardless of sales. The explanation the data rejects is that free tiers make tools acquisition targets: acquisition rates are flat across pricing models.

Are freemium AI tools more likely to be acquired? No. The freemium acquisition rate is 5.7% and the paid rate is 5.9%; every pricing model sits between 4.0% and 5.9%. Whatever acquirers are buying, it does not sort by pricing model.

Is pricing model a good way to judge whether an AI tool will survive? It is a weak signal. Category predicts survival about four times more strongly — a roughly sevenfold spread across categories against a 1.7× spread across pricing models. Check the category, the company's operating history, named customers, and whether it runs its own model before you weigh how it charges.

How was the AI tool shutdown rate measured? From the directory's own catalogue of published listings on 13 September 2026, counting confirmed shutdowns recorded in the AI Graveyard against all tools with that pricing model. Acquisitions are counted separately with a flag for whether the product still operates. It is a census of listed tools, not a sample of all AI companies, and the figures are frozen to their date.

Where to go next

The three research pages this draws on carry the full datasets: the AI tool shutdown census, failure rates by category, and what happens to acquired AI tools. For why a 5% figure and a 95% figure can both be correct, read why every source quotes a different AI failure rate. The AI Graveyard is the running list behind all of it, and what actually kills AI tools is the long-form read on the causes.

— The ToolDirectory.AI editorial team