Which models we track — and why some are missing
One test decides it: does anyone publish a per-token price for this model?
The ruleone test, applied the same way to everyone
A model gets a row when someone publishes a per-token price for it — the maker on its own rate card, or a hosting provider on theirs — and we can read that page directly and link you to it. That is the whole test. Today it yields 220 models from 32 providers, of which 192 are on sale right now.
This site is a price tracker. A row with no price cannot show you a price, and a number we cannot point at a published page is a number you should not trust. So the question we ask is not “is this model good?” but “can a reader actually buy tokens for it, at a rate someone has put in writing?”
What counts as a published price
Any of these gets a model listed:
- The maker's own rate card, pricing page, or API pricing docs.
- A hosting provider's rate card that names the model and prints its per-token rate — for open-weight models, this is usually the only price that exists, and it's the one you would actually pay.
- A published price in another currency, converted at a rate we disclose on the model's page.
These do not:
- A launch post, benchmark chart, or model card with no rates on it.
- Released weights with nobody hosting them for sale yet.
- A waitlist, a private preview, or a “contact sales” page that quotes no numbers.
- A third party's estimate of what a model “probably” costs. We'd rather have no row than a guessed one.
What never decides it
Benchmark scores don't. How large or famous the lab is doesn't. What country it's in doesn't. Neither does our own opinion of the model.
That's why the list can look lopsided at first glance. A little-known lab with a single model and a real rate card is here, because you can buy its tokens today. A widely praised frontier model that no one sells per token is not here, however well it scores — there is no price to publish, and inventing one would break the only promise this site makes.
A model being missing is not a verdict on its quality. It is a statement about whether anyone is selling it yet.
Exclusions are not permanent
Every model we've looked at and left out is written down: what we decided, why, and the date we decided it. Nothing is dropped quietly. Most of those records also state, in advance, the thing that would prove them wrong — usually some version of “somebody starts selling this per token.”
Those records are re-tested automatically every time we publish, several times a day. Two things happen. First, the stated reason is re-checked against current data, so an exclusion whose grounds have quietly evaporated fails loudly instead of ageing into a habit. Second, where the stated condition is one a machine can go and look at, we re-read the live sources on each publish: the public model catalog we watch, and the hosting providers' own rate cards. The moment either prints a paid per-token rate for a model we'd set aside, the exclusion stops holding and that model comes back for a fresh decision — which, for anything that has just been priced, normally means it gets added.
That live re-read doesn't yet cover every past decision — extending it is ongoing work, and the records without one are re-examined when we next sweep that vendor or model family. What is true of all of them is that the reasoning is on the record and dated, so a “not yet” can be revisited rather than re-argued from scratch.
Priced, and still not listed
Occasionally a price exists and we still hold the row back for a while. That happens when we can't yet tell what the number means: whether it's a launch promotion or the standing rate, which units it's quoted in, whether the input and output sides have been read off the same table, or whether two pages of the same vendor disagree. We resolve those against the vendor's own page before publishing rather than after. Omitting a model costs you one row; publishing a wrong price costs you the reason to believe the other 220.
Think something's missing?
Tell us and we'll check it — the fastest way to get a model listed is a link to the page where its price is published. Send it here. You can also see everything added recently on the new models page, browse the full model index, or read how each price is verified once it's in.