GPT-6 Astra takes first place on Code Arena's web development leaderboard
The board measures what voters prefer looking at, not whether the code is correct — a distinction worth holding on to when reading a 35-point gap.
Benchmark
Code Arena: WebDev
Human preference between two models asked to build the same front-end web application, including agentic coding workflows that require multi-step reasoning and tool use. Results are aggregated into an Elo-style rating.
- Maintainer
- Arena.ai
- Score
- Elo
- Last checked
- September 7, 2026
Leaderboard
No verified results recorded yet.
Methodology
Visitors submit a web development task, receive two anonymous attempts and vote for the one they prefer. Model identities are revealed after the vote. Ratings move continuously as votes arrive, and Arena.ai also publishes category-level breakdowns such as gaming, data analytics and full stack.
Important limitations
The rating measures what voters prefer to look at and interact with, not whether the code is correct, secure or maintainable. Voter populations are self-selected, prompts are not controlled, and a score is only comparable with others captured on the same date.