Arena announced on November 16, 2026, that Alibaba’s Qwen4-Max debuted at No. 6 on two web-development leaderboards: Code Arena: WebDev and Image-to-WebDev. Both scores are preliminary, and both are higher than those of Qwen3.8-Max (0902). Two matching sixth-place debuts make a tidy entrance. SourceSource
Here’s Arena’s reported scorecard:
- Code Arena: WebDev: Qwen4-Max scored 1,750, up 74 points from Qwen3.8-Max (0902)’s 1,676. Source
- Image-to-WebDev: Qwen4-Max scored 1,674, up 35 points from its predecessor’s 1,639. Source
On Image-to-WebDev specifically, that puts Qwen4-Max above Claude Opus 5 (Max), listed seventh with 1,664 points; GPT-6 Sol (Max), eighth with 1,657; and Muse Spark 1.3 (Max), tenth with 1,637. Source
Arena’s announcement posts don’t provide vote counts or confidence intervals to show how decisive those score gaps are. SourceSource
Better scores, smaller benchmark bill
In separate testing, Artificial Analysis reports that Qwen4 Max, using xhigh reasoning effort, scored 43 on its Intelligence Index, versus 34 for Qwen3.8 Max at xhigh and 46 for Kimi K3.1 Pro at max. Those are Intelligence Index scores, not Arena points. Source
Artificial Analysis also reports a cost of $3.27 per Intelligence Index task, down 67% from its predecessor’s $10.02. It attributes the reduction to lower pricing and lower token usage, including approximately 39% fewer output tokens per task. That’s a smaller bill for this evaluation workload, not a promise to knock two-thirds off everyone’s API spending. Source
On Terminal-Bench 4.0, Artificial Analysis reports Qwen4 Max at 48%, behind Kimi K3.1 Pro’s 67% but ahead of Tencent Hy4’s 37%. Source
A good website still needs to sell the ramen
In Arena’s first-impressions walkthrough, its AI Capability Lead, Peter Gostev, assessed Qwen4-Max as competent at straightforward websites while finding a substantial gap in more ambitious creative work. He recorded the qualitative walkthrough before he knew Qwen4-Max’s pricing or Arena ranking. Source
In one ramen-site comparison, he preferred GPT’s more direct route to ordering over Qwen’s scrolling-heavy approach. Attractive layouts are nice; letting someone buy dinner also helps. Source
Comments