What we never sell
We don't sell rankings. There is no way to pay for a model to move up the leaderboard — not for labs, not for partners, not for you. Every position comes from blind duels: two clips, same prompt, names hidden until after the vote.
XP can't be bought either. It only comes from judging, and it's never paid out as money.
Premium fights
Open models (Wan 2.2, LTX-2.3) render free in our own queue. If you want Veo, Kling or Seedance in the fight, you can render them here from a prepaid balance — or go to the maker's site; both are fine with us.
The price is the provider's cost plus a flat 25%, for every model. We never earn more from one model than another, so we have no reason to nudge you toward one. The model picker is ordered by leaderboard rank, not by price.
Paid fights earn no XP and count like any other clip if you send them to the arena: blind, after the same content review. If a lab or provider ever gives us a discount, it's listed here.
Affiliate links
When you click Try … on …, a link can be a referral link: if you sign up there, the site may pay us a fee. Such a link always carries a small Ad label.
A referral fee never affects rank. The leaderboard is computed from votes before any link is attached. Models without a program still get a plain link to their maker, so you can always try every model we rank.
Sponsored showcases
A lab or tool can pay for a showcase comparison — for example, a set of prompts that highlights what their model is good at. These are always labelled Sponsored on the card and the page.
Votes on sponsored comparisons are counted like any other, and the sponsor's model sits next to its rivals, rendered from the same prompts.
Which means the sponsor's model can lose — even on its own showcase. We publish the result either way and never hide a sponsored fight because of how the votes went. The truth can be brutal. 😅
Blind tests for labs
AI labs can pay us to blind-test a model before release. It appears in the arena under a code name, and real people vote on it exactly as they do on public models.
Results are only published with the lab's consent — but the votes stay real. We never adjust, filter or hand-pick them, and a model's public rank after release starts from the same duels.
How XP works
You earn XP for work that makes the leaderboard better:
- Pending first. New XP is pending for up to 24 hours while we check it.
- Checked. Accurate work is confirmed in full. Partly accurate work earns less (e.g. ×0.5); careless or made-up answers earn nothing.
- Accuracy. Your accuracy (High, Medium, Low) is how often the judge agrees with you. Only high-accuracy contributors appear on the weekly board.
- Daily cap. At most 500 XP a day from tasks; after 20 tasks each one earns a little less.
- Guests. You can vote without an account. Guest votes count on the leaderboard at a quarter weight, up to 20 a day; signed-up votes count in full. Up to 100 guest XP carries over when you sign up.
Levels unlock renders in the free queue (open models, our GPU, result by email, small watermark):
XP has no cash value and can't be bought.
Profiles show how much someone judges — duels, streaks, how often they agree with the crowd, their taste. They never show which clip you picked in a duel. Only you can see your own votes, under “My votes” on your profile. Not other users, not labs, not sponsors.
- It makes buying votes pointless. If votes were public, anyone could pay people to vote for a model and check the votes were delivered. When nobody can prove how they voted, a paid vote can't be verified — so it isn't worth paying for.
- It keeps the arena blind. If you could see that a top contributor always picks one model, you'd start voting like them instead of trusting your own eyes. Each vote should be one person's honest look at two clips.
- It protects you. Your picks say a lot about your taste. You might not want employers, labs or fans of a model to read through every duel you judged — and nobody should get grief for preferring the “wrong” model.
- It keeps our checks working. We spot vote rings by looking for patterns across many accounts. Public votes would show cheaters exactly which patterns get caught.
You can't make your votes public, even if you want to — that's what makes the promise hold for everyone. What we publish instead: totals per model, the methodology, and how many suspicious votes we removed. Your own votes are included when you export your data.
Your data
Your votes feed the public leaderboard. Flags, descriptions and votes also go into anonymised research datasets about where video models fail — no names, emails or handles attached.
Don't want your work in the datasets? Opt out any time in settings. Your votes still count on the leaderboard.