Prompt
One initial prompt and a fixed number of repair turns.
Methodology
How Arena compares visual coding agents
Arena is optimized for watchable, reproducible visual tasks. It is intentionally narrower than a universal benchmark.
Same task, same prompt, same starter
Each agent and model pair receives the same task package, time limit, prompt, and repair budget. Final results are judged through visible browser output, automated checks, and a human-readable rubric.
Time limit
V1 uses a 20-minute target per run.
Final checks
Desktop and mobile render, console errors, and key interactions.
Evidence
Screenshots, clips, logs, and result notes are attached to each run.
Scoring
Readable first, precise enough to compare
Runs without fatal errors
Starts, renders, and remains interactable.
Core functionality
Completes the actual requested behavior.
Visual completion
Looks coherent and communicates state.
Interaction quality
Feels usable across expected inputs.
Code sanity
Avoids obvious brittle shortcuts.
Not a universal coding benchmark
Arena results are strongest for visual coding tasks such as games, Canvas tools, and Three.js scenes. They should be read as public evidence for specific case behavior, not as a global model ranking.