The result
GPT-6 Astra (Max) is top of Arena.ai’s Code Arena: WebDev leaderboard with 1,797 points, ahead of Claude Fable 5.1 (Max) on 1,762 and Claude Opus 5 (Max) on 1,688. Alibaba’s Qwen3.8-Max-0902 sits fourth on 1,686 and Moonshot’s Kimi K3 (Max) fifth on 1,674. The board was last updated on 5 September and reflects 650,961 votes across 126 models.
OpenAI released Astra on 3 September, which makes this a placement earned in two days of voting.
How the board works
Code Arena is a crowdsourced Elo-style rating. A visitor submits a web development task, receives two anonymous attempts, and votes for the one they prefer; model identities are revealed after the vote. Arena says the tasks include agentic coding workflows requiring multi-step reasoning and tool use, as CryptoBriefing summarised.
That design has a real strength and a real weakness, and both matter here.
What 35 points does and does not mean
The strength is that the tasks are not a fixed dataset a lab can train against. Nobody can memorise the eval, because the eval is whatever a stranger typed in this morning.
The weakness is what a vote actually measures. A voter looks at two rendered results, briefly, and picks one. That rewards output that looks finished: complete layouts, plausible styling, confident structure. It does not reward code that is correct under conditions the voter did not test, secure against inputs the voter did not supply, or maintainable by someone else in six months. A 35-point Elo gap on a preference board is a statement about which output people liked better, and stops there.
It is also worth noting that a rating only means something against ratings captured on the same date. Boards like this move continuously as votes arrive, and the gap between first and second on 5 September is not a durable fact about either model.
The pricing angle
Both of the top two models are priced identically, at $10 per million input tokens and $50 per million output tokens, with one-million-token context windows, according to CryptoBriefing’s account of the leaderboard. That is the part of this result with commercial consequences: on a board that many buyers actually consult, the new leader did not arrive with a price premium.
Where the Chinese models sit
Fourth and fifth place going to Qwen3.8-Max and Kimi K3 is the quieter finding. Both trail the leaders, but by less than the gap between first and third, on a board voted by whoever showed up rather than by a lab’s own evaluation harness.
What to watch
Whether Astra holds the position as the vote count grows — early placements on preference boards are noisy, and a two-day-old model has a novelty advantage that decays. The more informative signal will be where these five models sit on the same board in a month, and whether anything with published weights closes the remaining gap.
