Code Arena
An AI-assisted programming evaluation platform under Arena.ai. It has introduced a FullStack feature supporting the construction of complex applications that depend on databases or backends, and provides a community-voting-driven leaderboard.
Timeline (last 90 days)
According to La Marina's Code Arena Web development leaderboard, GPT-6.1 Terra Max scores 1671, ranking 14th, 151 points higher than GPT-5.6 Terra.
The top spot on the Code Arena leaderboard is Claude Opus 5.5 Max with a score of 1809.
Gemini 4 Argon ranks only eighth in the Code Arena Web Dev category on Code Arena, slightly above QWEN 3.8.
Claude Opus 5.5 topped the Web Dev leaderboard on Code Arena with a score of 1818, an improvement of 126 points over Opus 5, and is approximately 60% cheaper than the second-place GPT-6.
Arena mechanisms based on human blind evaluation and real-task benchmarks can largely circumvent the overfitting (teaching to the test) problem of static benchmarks.
GLM-5.2 ranked second globally in Code Arena with a score of 1595.
Code Arena is a competitive arena focused on front-end code generation that uses blind user voting to produce Elo scores.
Qwen3.8-Max-0902 tops the Pareto frontier leaderboard (at the $5/MTok price tier), leading Claude Opus 3.5 by 3 points.
Qwen3.8-Max-0902 ranked first overall on the Code Arena comprehensive leaderboard with a score of 1691.
GLM 5.3 has entered the web development frontier leaderboard on Code Arena.
All Facts (5)
Code Arena uses a blind-test battle mechanism, with results aggregated into Elo scores calculated based on the Bradley-Terry model.
60%UnverifiedAccording to La Marina's Code Arena Web development leaderboard, GPT-6.1 Terra Max scores 1671, ranking 14th, 151 points higher than GPT-5.6 Terra.
50%UnverifiedThe top spot on the Code Arena leaderboard is Claude Opus 5.5 Max with a score of 1809.
50%UnverifiedCode Arena is a competitive arena focused on front-end code generation that uses blind user voting to produce Elo scores.
50%UnverifiedArena mechanisms based on human blind evaluation and real-task benchmarks can largely circumvent the overfitting (teaching to the test) problem of static benchmarks.
50%