methodology and disclosures

two kinds of entry, and the page always says which

a listing either carries a verdict in our own words or it carries the vendor's description. the first kind came from primary-source research: we read the vendor's own pricing page, the actual licence file where one exists, and the product documentation, and wrote a line about what we found. the second kind is there so a category shows the whole field rather than only what we have got to, and its page says outright that we have not tested it.

neither kind is a laboratory benchmark, and we don't want to imply otherwise — plenty of review sites publish precise-looking accuracy percentages no reader can reproduce. where a category has credible independent measurement, such as a public leaderboard with a large vote pool, we use it and say so. otherwise a verdict is a considered opinion informed by the documents above, and it reads like one.

we read the licence, not the launch blog

this is the part of the process that has caught the most real errors, and it is unglamorous: we open the LICENSE file rather than trusting the announcement.

in 2026 those two things disagree constantly. black forest labs' own flux.2 post reads as though the downloadable model is apache 2.0; the licence on hugging face is non-commercial. fal's model pages describe ltx-2 as apache 2.0; the actual licence carries a $10m revenue gate and an anti-competition clause. ideogram 4 is widely reported as commercially usable open weights; the model cards say non-commercial. tencent's hunyuan licence excludes the eu, the uk and south korea outright, which almost no coverage mentions.

if you take one thing from this site, take that habit rather than our opinion.

why there is no score

there used to be one, out of 100, and it did more harm than good: a number invites comparison between things that were never measured the same way, and an 85 for an image model and an 85 for a verification tool are not the same claim.

what replaced it is a sentence. where we researched a product, its page says what we thought and, more usefully, what the weakness is. capability, operational reality — reliability, latency, support, how bad the failure modes are — cost at a realistic volume rather than at the headline tier, and viability all still go into that sentence. a product with a published shutdown date, a licence you cannot ship on, or no api to build against gets said so plainly.

the order of a category page is not a quality judgement either: paid slots first, then whatever readers have upvoted. that is stated on the page so nobody mistakes position for merit.

what we don't claim

we don't publish measured figures we didn't produce. if you see a percentage on this site, it comes from a named source — a vendor's own claim, a public leaderboard, or a licence or pricing document — and the entry says which.

where we couldn't verify something, the entry says so rather than filling the gap with a plausible number. there are several of those on this site by design: reve's per-image price for version 2.1 isn't published anywhere we could reach, minimax sells video generation in opaque 'points' with no per-second rate, and midjourney blocks automated access to its own pricing page.

gaps are more useful than confident guesses, and they're honest about where a reader should go and check for themselves.

what happened to the rankings

this site used to publish ranked lists — best 10 of this, best 15 of that — at their own urls, and for a while also at /blog. they are gone, and every one of those urls now redirects to the directory page for the same category or the listing for the same product.

the research did not go anywhere: it is what the verdicts on those listings are made of, and the licence and pricing work described above is the same work. what went is the ordered list and the score, because ranking a field that changes weekly means publishing an order that is stale before the page is.

how this site makes money

directory listings. a free listing costs nothing and is reviewed by hand; a featured slot is $50 one-time, publishes immediately, and pins the listing above the free ones on its category page. every featured listing is labelled as such wherever it appears, and its outbound links carry rel="sponsored nofollow".

also advertising, and affiliate links on some outbound vendor links, where we may earn a commission if you sign up after clicking through.

none of it reaches a verdict. a listing fee cannot get one written, cannot soften one already written, cannot remove or demote a competitor's entry, and has never done any of those things. that line is the only reason anything editorial here is worth reading, so it is the one thing on this page that is not negotiable — if it ever changes, this page changes with it first.

where we operate the tool ourselves, the entry says so on its face rather than in a footer. currently that applies to the free email checker at verifier.org.

vendors don't get a preview

no vendor sees a verdict before it publishes, and no vendor can make a re-check a condition of anything.

we correct factual errors quickly, from anyone. if a price, a limit, a licence term or a feature claim is wrong, tell us and we'll fix it — this category moves fast enough that some of it will be wrong by the time you read it.

when an entry is re-checked

when something material happens — a major version, a pricing change, a licence change, an acquisition, or a shutdown announcement — not on a calendar.

prices move faster than anything else here, which is why a listing quotes one only where we read it off the vendor's own page and can say when. everywhere else the page says 'not stated' rather than guessing, and sends you to the vendor to look.