People send an image with a prompt, compare answers from two anonymous vision-language models, and vote for the better one. Votes become Bradley-Terry ratings.
Arena ranks models as the orchestrator in real Agent Mode sessions on arena.ai. Scores use behavioral signals such as file downloads, disapproval events, retries and steerability.
Users upload a screenshot, wireframe or mockup, two anonymous models build the website in Code Arena, and users vote for the better result. Votes become ratings with bootstrapped confidence intervals.
People upload their own PDFs with a question, compare answers from two anonymous models, and vote for the better one. Votes become Bradley-Terry ratings.
People compare answers from two anonymous search-augmented models on their own questions and vote for the better one. Votes become Bradley-Terry ratings.
Arena-Hard-Auto500 hard arena prompts judged by an LLM against a baseline. Cheap proxy for arena rank.
Overall chat quality
Apr 2024
97.1%
Fading
Arena-Hard v2.0Models answer 500 hard real user prompts from Chatbot Arena and 250 creative writing prompts. GPT-4.1 or Gemini 2.5 judges each answer against a baseline answer from o3-mini, and the score is the win rate.
Overall chat quality
Apr 2025
90.2%
Fading
Alignment Index (Arena)Arena ranks agent models on failure signals mined from real Agent Arena sessions, such as unauthorized actions, false attribution and deceptive completion.
Deception and alignment
Oct 2026
87.9
Not in cards
VisionArena-BenchModels answer visual questions from Arena users. A vision language model judges their answers to approximate human preference rankings.
General visual questions
Dec 2024
88
Not in cards
Arena-T2I HardModels generate images from detailed user prompts. A judge checks whether the images satisfy individual requirements and their dependencies.
Image generation and editing
Jun 2026
0.85
Not in cards
Preference Proxy EvaluationsReward models and judges rank responses. The benchmark compares their decisions against user preferences and verifiable answers.
Overall chat quality
Oct 2024
68.5%
Not in cards
Arena Search FactualityArena combines user preferences with independent checks of verifiable claims in model responses.
Long-form factuality
Jul 2026
1,218
Not in cards
ArenaWebdevCorrectnessJudges inspect generated applications and predict whether their features work. Human labels provide the reference answers.
Web and app front ends
Oct 2026
81.6%
Not in cards
ArenaWebdevPreferencesJudges compare pairs of generated applications and predict which application the user preferred. The score measures agreement with user votes.
Web and app front ends
Oct 2026
69.1%
Not in cards
On Evaluating User Simulators for Interactive AgentsModels play the user in replays of 351 real sessions. The evaluation compares simulated behavior with observed behavior and measures agreement with human rankings of nine agents.
Multi-turn conversation
Nov 2026
Not in cards
Arenas without a leaderboard here
Name
Kind
Released
Best score
Status
BiomedArenaResearchers compare anonymous answers to biomedical research questions and vote for the response they prefer.
Expert-written questions
Aug 2025
Not in cards
RedTeam ArenaPlayers get 60 seconds to make a model say a target word. An extended Elo ranks players, models and target prompts, with models ranked by resistance.
Jailbreaks and prompt injection
Sep 2024
Not in cards
RepoChat ArenaUsers give a GitHub repository link and a question. A retriever model picks files, two answer models reply, and users vote for the better answer.