Text
Head-to-head ratings for general-purpose chat and reasoning, with style control applied so answers aren't rewarded just for being long or prettily formatted.
Leader
claude-fable-5
1,507.62 · anthropic
5 ranked models
Community rankings
AI Model Leaderboards
Blind head-to-head votes from Arena. We publish the source data and never adjust rankings for sponsors.
WebDev
Which models write the best working front-end code, judged by people comparing two real generated apps side by side.
Leader
claude-opus-5-max
1,711.88 · anthropic
10 ranked models
Text to Image
Image generators ranked by blind human preference — two images from the same prompt, people pick the better one.
Leader
gpt-image-2 (medium)
1,384.8 · openai
5 ranked models
Image Edit
How well models edit an existing image to match an instruction — inpainting, restyling, object removal.
Leader
gpt-image-2 (medium)
1,462.74 · openai
8 ranked models
Text to Video
Video generators ranked on prompt-to-clip quality — motion coherence, prompt adherence and overall watchability.
Leader
gemini-omni-flash
1,527.49 · google
10 ranked models
Image to Video
Animating a still image into video — the workflow most creators actually use for controlled results.
Leader
dreamina-seedance-2.0-720p
1,473.66 · bytedance
6 ranked models
Video Edit
Editing existing footage from an instruction. The newest and smallest board, so treat the ordering as provisional.
Leader
dreamina-seedance-2.0-720p
1,377.13 · bytedance
6 ranked models
Vision
Understanding images rather than making them — reading charts, screenshots, diagrams and photos.
Leader
claude-fable-5
1,334.18 · anthropic
6 ranked models
Document
Working with long documents and PDFs — extraction, summarizing and answering questions grounded in a file.
Leader
claude-opus-5-high
1,520.11 · anthropic
6 ranked models
Search
Web-grounded answering — how well a model finds current information and cites it accurately.
Leader
claude-opus-4-6-search
1,252.82 · anthropic
8 ranked models
Attribution and methodology
Rankings are from Arena, based on community blind comparisons. Source: Arena Leaderboard Dataset, licensed CC BY 4.0. Close confidence intervals should be treated as ties.
Last synced (UTC): July 30, 2026 at 12:00 AM