How the Fantasy Arena ranks models
Every ranking comes from people choosing between two outputs for the same fantasy prompt without knowing which model made which. This page explains what counts, how scores are calculated, how sure they are, and how to use the data.
The vote
- Same prompt. Two models receive the same fantasy or tabletop prompt through CharGen’s normal generation paths, for one task: for example Character, Monster, Battlemap or Handout for images.
- Names hidden. Voters see Left and Right. The sides are randomised, and the model names are only revealed after a vote is recorded.
- Who votes. Signed-in CharGen users, one vote per user per match. Left better, right better and tie are counted. “Both poor” is kept as feedback and skips are ignored; neither moves a score.
The score
Scores are a Bradley-Terry fit over every counted vote on a board, on a familiar scale centred on 1,000. A gentle pull toward 1,000 (the equivalent of three wins and three losses against an average model) stops a model with a handful of lucky wins from topping the board. A tie counts as half a win to each side. Each task board is fitted on its own votes.
Each score carries a 95% range, from 200 resamples of the votes. The rank range shows the best and worst place the ranges allow: “rank 3, could be 2 to 6” means the order between those models is not settled yet. With the vote counts this board has today, overlapping ranges are normal, and the board shows them rather than pretending to precision.
When a model is ranked
- In evaluation
- Under 20 counted votes. Listed below the table, not ranked.
- Preliminary
- 20 to 49 counted votes. Ranked, with a badge: the score can still move a lot.
- Ranked
- 50 or more counted votes.
Which matches you are shown
Each ballot starts from up to 100 ready matches you have not rated. One ballot in four is picked at random from them. Otherwise matches are grouped by the less-voted of their two models, and a group counts four times when that model is in evaluation, twice when it is preliminary and once when it is ranked; a group is drawn by weight, then a match at random inside it. The weight belongs to the model, not to each of its matches, so no model can take more than its share of ballots. Choosing pairs this way does not bias the scores, because who wins a pairing does not depend on how the pairing was chosen.
New models enter through matches people create with Run a test. CharGen also pays for matches from a house account to bring new and low-vote models into the pool; that account never votes.
Snapshots and data
On the 1st of each month, every board (overall and each task) is frozen as it stood at the end of the month: the score, its range, the rank range, vote counts and status. Snapshots never change once published and are available as JSON and CSV under CC BY 4.0. The live boards keep updating between snapshots.
How to cite
Cite a monthly snapshot rather than the live board, so the numbers you quote stay checkable:
CharGen Fantasy Arena, Image board, October 2026 snapshot. char-gen.com/ai-model-leaderboard/image
@misc{chargen_fantasy_arena,
title = {CharGen Fantasy Arena},
author = {CharGen},
year = {2026},
url = {https://char-gen.com/ai-model-leaderboard},
note = {Monthly snapshot, CC BY 4.0}
}What a score is not
Scores are preferences from tabletop players and makers on fantasy prompts. They are not objective quality measures, and a model that ranks well here may rank differently on other kinds of work.