Qwen3.8-Max-0902 Tops WebDev at a Quarter of Opus 5's Price, API-Only
Alibaba's Qwen3.8-Max-0902 took first place on the Code Arena WebDev leaderboard at 1,691 points, edging out Claude Opus 5 Max at 1,687, while charging roughly a quarter of the price. The catch: the 0902 snapshot is API-only. Only the base model has downloadable weights, so treat the "open-weight" framing in early coverage with skepticism.
What the numbers actually say
The WebDev win is real but narrow: three points on one leaderboard, in a category that favors front-end work. On backend and agentic coding suites, Claude Opus 5 Max still leads comfortably: TerminalBench 3.0 (42.7 vs 29.0), DeepSWE 1.1 (73.6 vs 69.3), ProgramBench (41.5 vs 28.0). Across 15 scored benchmarks, Opus 5 leads in 8 and Qwen in 7. This is a split decision, not a sweep.
The pricing gap holds up better than the leaderboard claim. At a 1:3 input-to-output mix, Qwen3.8-Max-0902 ($2 input, $6 output) blends to about $5 per million tokens, against about $20 for Claude Opus 5 ($5 input, $25 output): a 4x reduction for competitive output on web development tasks.
Models mentioned: Qwen3.8 Max · Claude Opus 5 · Kimi K3
See also: Best Coding Models · Best Cheap Models
Who should care
- Web development on API budgets: Qwen3.8-Max-0902 is the value pick where it leads, at about a quarter of the Opus 5 price.
- Backend and agentic workflows: Opus 5 keeps the edge on TerminalBench, DeepSWE, and ProgramBench. Benchmark your own workload before switching.
- Self-hosters: wait. The 0902 snapshot has no published weights; the downloadable base model is a different artifact under a custom license.
Takeaway
Qwen3.8-Max-0902 resets the price floor for top-tier web coding output, but the three-point lead is churn-prone and Alibaba's in-house benchmark claims still await independent replication. Good deal, narrow crown, not open weights.