We list 268 language models. That number describes our list, not the market.
Four different counts come out of the same catalogue, from 291 down to 65, and none of them is the number of AI models in the world. Where each count comes from, and what it can and cannot tell you.
On 14 November 2026, our catalogue held 291 language-model entries. Apply our variant filter and 268 remain. Those are fair numbers to quote about the AI Resource Hub. They are poor numbers to quote about AI.
A catalogue records what its collectors found and what its editors added. A model that never reached either route is absent, and absence tells you nothing about whether the model exists, works well or is worth using.
Four counts from one list
The figures in our recent articles all come from the same frozen copy of the catalogue, but they do not all count the same thing. Here are the four.
| Count | Entries | What it counts |
|---|---|---|
| Language-model entries | 291 | Every record in the language-model category |
| After variant filtering | 268 | Drops 23 entries whose id ends in a date stamp or in "-fast" |
| With a date and a context window | 252 | The records our context-window charts can use. 16 of the 268 are missing a release date, a context window, or both. |
| Active and priced | 65 | The records our price comparisons use: marked active, with input and output prices above zero |
Four counts from one catalogue
The last two bars are why a context-window chart and a price chart from us can describe different sets of models, even though both say "our catalogue". The article on context windows uses the 252. The one onthe price spread uses the 65.
How a model gets into the list
There are two routes. Some models are entered by hand from lab announcements and documentation. Others are added automatically when OpenRouter lists a paid model from a provider we watch. Those start with the status "tracking" and a note that they are awaiting verification against the lab's own page.
The automatic route has limits that come with it. OpenRouter's documentation (read on 14 November 2026) says its models list returns text-output models only unless you ask for other types. Our collector also skips free variants, skips routing variants, and only auto-tracks a model if its provider is one we recognise. A model that is free only, from an unfamiliar lab, or not a text model, would not arrive that way.
The hand-entered route has gaps too, since it depends on what we noticed and added. We cannot say how many models both routes miss, because that would need an independent list to compare against, and we do not have one.
What follows for anyone using the list
- A missing model is not a verdict. It may be unavailable, unsuitable or simply unlisted. For a model you care about, the lab's own documentation is the source.
- A count is only useful with its denominator. "65 active, priced models" is a statement about a defined selection. "291 models" is a statement about our records.
- Some charts cannot be drawn yet. The open-source flag on automatically added models is unreliable, as ourtransparency page says, so we do not publish an open-versus-closed breakdown.
None of this makes the catalogue useless. It is a dated list with stated rules, and it is good for finding and comparing the models it holds. It is a poor way to count how many models exist, and the numbers in this article are a reason to read any "number of AI models" headline, ours included, with its denominator beside it.