How ELO works here
1
Everyone starts level
A new model starts at 1000.
2
Wins move points
When a model wins a vote it gains points and its opponent loses points. Beating a higher-rated model is worth more than beating a lower-rated one.
3
Upsets matter most
Losing to a much lower-rated model costs more than losing to a strong one. Over many battles, scores settle where the community’s preferences put them.
ELO adjusts for who a model beat, which a raw win rate can’t. A model that only ever met weak opponents can have a high win rate and a modest ELO.
The math
The math
For models with ratings RA and RB, the expected score of A isEA = 1 / (1 + 10(RB − RA) / 400)After a vote, A’s new rating is RA + K × (SA − EA), where SA is 1 for a win, 0 for a loss and 0.5 for a tie, and K sets how fast ratings move. B is updated the same way, so a win against a stronger opponent (low EA) earns more points.
Reading the board
eloRank
Position by ELO, 1 is best. Models with exactly the same score share a rank.
eloScore
The current rating. Differences matter more than absolute values.
totalWins · totalLosses · totalTies
Vote outcomes involving the model.
totalResponses
How many comparisons the model has appeared in. More battles means a more settled score.
winRate
Wins as a percentage of battles (0 to 100).
Live leaderboard
Straight from the public API, refreshed when you load the page. Filter by provider or by how many battles a model has played.Get the data
The leaderboard is public. No account or token is needed:Caveats
Small samples are noisy
Small samples are noisy
A model with a handful of battles can sit high or low by chance. Look at the battle count before drawing conclusions.
The prompts come from the community
The prompts come from the community
Rankings reflect what Arena users ask. If most prompts are about code, the board leans toward coding ability.
One board for all tasks
One board for all tasks
There is a single overall ranking today. Per-category boards are on the roadmap.