Blind Side-by-Side Battles
A user submits an open-ended prompt—ranging from complex full-stack web apps to photorealistic video synthesis. Two anonymous models generate outputs simultaneously.
No matching models
"Preliminary" indicates fewer than standard evaluation battles completed; confidence intervals (±) narrow as additional votes are recorded. Click column headers to sort.
EVALUATION PROTOCOL
How the Rankings Work
Unlike static academic benchmarks that models can memorize, our leaderboard reflects real-world human preference through anonymous pairwise battles.
Human Preference Voting
Before model identities are revealed, the user evaluates both responses for accuracy, instruction following, aesthetics, and coherence, then casts a vote for the superior result.
Bradley-Terry Elo Calibration
Millions of pairwise battle outcomes are fitted using the Bradley-Terry statistical model to compute stable Elo ratings and 95% bootstrap confidence intervals (±).
KNOWLEDGE BASE
Frequently Asked Questions
What does the Elo Score and Confidence Interval (±) mean?
Similar to competitive chess, an Elo score measures relative skill across head-to-head matchups. A 100-point Elo advantage translates to roughly a 64% expected win rate against the lower-ranked model. The ± value shows the 95% confidence interval—models with more votes have tighter intervals.
How is the Code Arena evaluated compared to general Text?
The Code leaderboard specifically focuses on software engineering, debugging, and interactive WebDev tasks where models build functional HTML/CSS/JS interfaces and applications that human evaluators test directly in a sandboxed preview.
What does "Price $/M (in/out)" and "Context" represent?
For language and coding models, API pricing is listed in US Dollars per 1 million input tokens / 1 million output tokens. Context indicates the maximum token window the model can process in a single conversation.
How frequently is the leaderboard data updated?
We synchronize leaderboard snapshots whenever new major frontier models finish initial blind evaluation batches or when community vote counts cross significant statistical thresholds.