About
What ModelCap is, who runs it, and how it pays for itself
ModelCap is a live ranking board for AI models. It reads public leaderboards, model cards, price lists and provider catalogues on a fixed schedule, turns them into one auditable score per model, and publishes the result with its uncertainty. Nobody hand-places a model, nobody pays for a position, and no language model decides a number.
Who runs it
ModelCap is operated by Model Cap LLC, an independent company. The methodology explains how the ranking works, and the public dataset provides downloadable records. The application repository is private. Publishing the methodology and data does not mean the complete implementation or every historical input is available for independent replay.
ModelCap brings results from multiple leaderboards together so readers can compare capability evidence, cost and availability in one place. The active method determines the ranking statistic and publishes its version alongside the data. Capability estimates, evidence support and uncertainty are separate: a high position alone does not establish broad independent evidence.
How the board is made
Everything on the board is derived from named sources: Arena, the Artificial Analysis evaluation boards, ARC Prize, SWE-bench, BFCL, WildClawBench, publisher model cards on Hugging Face, and the OpenRouter, Vercel AI Gateway and Hugging Face catalogues for prices, context windows and availability. Each source is listed on the methodology page with its role and the weight it carries.
A data plane in AWS refreshes discovery about every five minutes and recomputes the whole board roughly every fifteen. If a refresh fails validation, the previous good board keeps serving; the site never publishes a half-computed rank. The scoring code is deterministic, version-stamped, and replayable: a sealed board can be regenerated from its inputs and must reproduce exactly.
- Published numbers from a publisher's own model card can locate a brand-new model on day one, but they are quarantined and never become measured evidence. Only independent boards do that.
- Downloads, likes, prices, provider counts and parameter counts are never quality signals. They are displayed, not scored.
- A model that cannot be tied to one exact catalogue identity gets no rank at all. We would rather show a gap than a guess.
The rules change when the evidence landscape changes, and every change is written up. The notes record what the board did on a given day and why a rule exists; the methodology is the current contract.
How it is funded
ModelCap plans to support the site with Google AdSense advertising, subject to approval. When enabled, ad units use fixed, labelled positions on a handful of page types, never inside the ranking table, a model's evidence panel, or the navigation.
No advertiser, publisher or provider can buy a position, a badge or a mention on the board. We do not accept payment to add, remove, rename or re-rank a model, and we do not run affiliate links on model or provider pages. If a page ever carries a paid placement outside the Google network, it is labelled as sponsored and it is not part of the ranking.
What ModelCap is not
- Not a benchmark. We do not run models ourselves. “Measured” means that independent public boards measured the exact catalogue product; “Modeled” means we inferred a position and say how.
- Not a review site. There are no star ratings, user reviews or editorial picks. A rank is a computed statistic with a stated confidence.
- Not advice. The prediction-market page shows public market prices about AI as context. Nothing on the site is investment, legal or purchasing advice.
Corrections
When a number is wrong, the fix is almost never to edit the number. It is to find the rule or the source that produced it and change that, so the same mistake cannot recur for the next model. If you think a model is misidentified, mispriced, missing, or ranked from evidence that does not belong to it, write to us with the model id and the source you are looking at. Details are on the contact page.
The published dataset, its licence, and the field definitions are on the public dataset page. The privacy notice is at /privacy.