We asked all 701 AI tools what they cost. About a third would not say.

Of every question people ask about an AI tool, the one they ask most is what it costs. We measured that directly last month: across the 45 most-voted listings here, <tool> pricing draws 66,140 searches a month — forty-eight times the volume of "is it worth it", and it exists for 43 of those 45 tools where "worth it" exists for 20.

So we asked all 701 listings the same question, on 10 August 2026: open the pricing page and see whether there is a number on it.

The answer

That is what the machine found: a currency figure on the pricing pages of 512 of 697. It is also, as it turns out, too generous — see the correction two sections down, which brings it to roughly 453, or 65%.

The 185 that showed no figure split into two quite different behaviours: 75 have built a pricing page and used it to ask for a conversation, and 110 have no pricing page at all.

The second group is the larger one, and mostly not a strategy. It is small tools, side projects and research releases that never wrote the page. Fifty of the 110 are listings we have labelled free, where there is genuinely nothing to price.

The part we nearly got wrong

Our first pass fetched each pricing page and looked for a currency symbol. It concluded that 61% publish a price and that 29 sites were unreachable.

Both numbers were wrong, and wrong in the tool's disfavour. Clipdrop's pricing page is 79KB of HTML containing no $ anywhere — not in the markup, not in a script payload. The prices arrive from an API after the page loads. To a fetch, a tool that renders its prices in JavaScript is indistinguishable from a tool that refuses to quote one, and those are opposite findings.

So we ran the negatives again through a real browser. 105 listings changed their answer. Twenty-five of the 29 "unreachable" sites were simply refusing our fetch and served a normal page to a browser.

Even that was not enough on the first try. A pass that waited for the document plus two and a half seconds put Recraft in the contact-sales column; its pricing page publishes seven figures from $0 to $211, and they had not arrived yet. Waiting for the network to go quiet instead moved it, and others like it. A third pass changed only 8 more listings out of 197, which is how we know the number has stopped moving.

If you see a study claiming most AI tools hide their pricing, check whether it rendered the pages. Ours would have overstated it by twelve points.

Where our own labels disagree with the page

Every listing here carries a pricing label. Set against what the pages actually show, one column stands out:

Fifty-one of the 107 listings we label "contact sales" appeared to publish a figure. So we read all 51 by hand, and that is where this article stopped being a machine's answer.

The correction: one figure in ten is not a price

A currency figure on a pricing page is not proof of a plan price. We knew that in the abstract. Reading the pages showed how it actually happens.

Of the 51, 48 gave us something to read, and 16 of those were not prices at all:

• Acrolinx quotes customers as "$30B Med Device Manufacturer" and "$8B FinServ Firm"
• DataRobot leads with "$60M ROI" and "$200M ROI" from case studies
• Duckie renders a product demo containing $49 / mo inside a mocked-up Stripe call
• One Word Domains lists .com domains at $72.40 — a marketplace's inventory, not its own price
• Others carry a funding round, an invoice in a screenshot, or a hackathon prize

Sixteen of forty-eight is a third, but that group is the worst case: "contact sales" listings are disproportionately enterprise sites, and enterprise marketing is made of large dollar figures that are not prices.

So we sampled the rest. Forty-five listings drawn at random from the other 461, read the same way: four were not prices — Groq's $650M fundraise banner, an example travel query mentioning "$49", a benchmark table of other models' API costs, and a promotional "get $25 of credit" offer. Four in forty-four with evidence, or 9%.

Applying both rates, roughly 453 of 697 — 65% — publish a price, and 244, or 35%, do not. Not 73% and 27%. The sample is small enough that the honest range on the larger group runs from about 58% to 69%.

We are not relabelling the catalogue from a script, for exactly the reason this section exists. Of the 32 that turned out to publish real prices, 29 have been relabelled from the evidence — eleven to freemium where a $0 tier sits beside paid ones, nine to subscription, nine to metered. Three were left alone because the three lines we captured did not settle whether a free tier exists, and guessing is how this goes wrong.

What this means if you are the one being searched for

If you make one of the 244: the people typing your product's name followed by "pricing" are the furthest down the funnel of anyone who will ever search for you. They have decided to evaluate you. A page that answers them with a contact form hands that search to whoever will answer it — a directory, a review site, a competitor's comparison page, a Reddit thread from 2024 quoting a price you no longer charge.

That is not an argument against enterprise pricing. It is an argument for putting a number on the page even when the real answer is "it depends" — a floor, a per-seat rate, a usage rate, anything a search result can quote back.

And if your prices are rendered in JavaScript, know that you are invisible to a large class of readers: not only our first pass, but plenty of scrapers, previews and summarisers that never run your scripts.

What we are not claiming

One check per listing per path, from one machine, on one day, with one browser. Three passes agreeing is evidence the instrument is stable, not that it is right.

The largest limitation is that all three passes used the same headless Chromium. A site that specifically blocks headless browsers looks exactly like a site with no prices, and we cannot separate those two from the outside. Four listings blocked us outright — Perplexity, Midjourney, Gemini and Magnific — and we say nothing about their pricing here.

We checked ten of the 185 that showed no figure by hand, at their real pricing URLs, with a longer wait than the census used. All ten came back with no figure, which is reassuring and is still only ten. The hand-reading went the other way — into the listings that did show a figure — because that is where a wrong answer flatters the tool, and those are the ones a directory should be hardest on.

"Shows a price" is a claim about the page, not about what you will pay. We did not attempt to read plans, tiers, or what any figure buys. The one number in this article that is a judgement rather than a measurement is the 51 mislabelled listings, and that one is about us.