Code Arena is a live evaluation platform that ranks AI models by watching them build actual software, rather than scoring them on a fixed set of test questions. Two anonymous models get the same open-ended coding prompt, build a working application inside a sandboxed environment, and a human judge picks the better result. Those votes accumulate into a running leaderboard, and the ranking has become one of the most closely watched signals for which large language models are genuinely good at software development, not just at answering questions about it.
How the Evaluation Works
Each round starts with a prompt — “build a expense tracker with charts,” for instance. The two competing models act as agents: they plan the task, then use structured tool calls such as create_file, edit_file, and run_command to write, execute, and revise code step by step, the same way a developer would work inside an editor and terminal. Every action is logged and the session stays inspectable afterward, so a reviewer can see not just the final app but how the model got there.
Once both models finish, a human evaluator compares the two builds side by side and votes on three dimensions: whether the app actually works, how pleasant it is to use, and how faithfully it matches what was asked for. Those pairwise votes feed an Elo rating system — the same style of formula chess uses to rate players — which converts thousands of head-to-head wins and losses into a single comparative score. Because any individual score is really an estimate, the leaderboard also shows a confidence range around each model’s position rather than pretending the ranking is exact.
From Chatbot Arena to Code Arena
The format descends from Chatbot Arena, the blind side-by-side voting site that LMArena — now simply called Arena — built in 2023 to rank general conversational responses. As AI labs shifted toward agentic coding tools that plan, use tools, and iterate rather than just autocomplete a function, the same team built a dedicated coding version, first as WebDev Arena and then, in a broader relaunch, as Code Arena. It now covers full-stack tasks — databases, authentication, third-party integrations — not just front-end prototypes.
Why It Matters
Older coding benchmarks mostly check whether generated code passes a fixed set of unit tests. That approach rewards models for pattern-matching known problems and says little about messier, real-world work: turning a vague request into a working product, recovering from an error, or making sensible design choices along the way. Code Arena’s agentic, human-judged format is an attempt to measure that instead, and its results now influence which models developers and companies choose to build with — the same role AI benchmarks have long played, just applied to a harder, more open-ended task.
It is not a perfect substitute for trying a model yourself. A model’s Code Arena rank reflects one specific style of task — building a self-contained app from scratch — and can diverge sharply from how it performs on debugging an existing codebase, writing infrastructure code, or general conversation. Rankings also shift quickly as labs release updated models, and a crowd of human voters, however large, still carries its own biases about what looks polished. Treat the leaderboard as a useful, evolving signal rather than a final verdict.
Readers curious to see it firsthand can browse the live rankings — and cast their own votes — at Code Arena’s WebDev leaderboard.
In the News
Code Arena’s rankings shift often enough to make news in their own right: Alibaba’s Qwen model recently topped the Code Arena leaderboard, overtaking Claude Opus 5 — a reminder of how quickly the top spot changes hands as labs ship new versions.