Alibaba's Qwen tops the web development leaderboard with 1,691 points: 17 points ahead of Kimi, output price less than a quarter of Claude's. On September 2, Alibaba's Tongyi Qianwen officially launched Qwen3.8-Max-0902, the latest iteration of Qwen3.8-Max. On the same day, the WebDev (web development) leaderboard of the overseas code arena Code Arena was updated:

💡 What You Will Learn

Alibaba's Qwen tops the web development leaderboard with 1,691 points: 17 points ahead of Kimi, output price less than a quarter of Claude's. On September 2, Alibaba's Tongyi Qianwen officially launch

📜 Table of Contents

• 1691 points, pushing both Claude and Kimi behind it
• The cheapest at the top, the highest-scoring among the cheap
• Not just building web pages: long tasks that run for five straight days
• Topping the web leaderboard, but coding isn't all-around yet

Alibaba's Qwen Tops the Web Development Leaderboard with 1691 Points: 17 Points Ahead of Kimi, Output Price Less Than a Quarter of Claude's

On September 2, Alibaba's Tongyi Qianwen officially launched Qwen3.8-Max-0902, the latest iteration of Qwen3.8-Max. On the same day, the overseas code arena Code Arena updated its WebDev (web development) leaderboard: the new version of Qwen shot to first place overall with 1691 points.

Code Arena is the AI coding blind-test leaderboard that overseas developers check most often, backed by the Arena evaluation system built by a Berkeley team, with over 600,000 real human votes accumulated in the web development track. Its rules are completely different from an exam: the prompts are real needs typed out casually by real users, undisclosed and unfixed, so model vendors can't optimize against the question bank in advance; two models anonymously submit finished web products, users vote on them in pairwise comparisons, and points are calculated using the Elo algorithm from esports ladders. Vendors can pick their own arena for the benchmarks they publish themselves, but they can't pick this leaderboard.

1691 points, pushing both Claude and Kimi behind it

The top few after the leaderboard update are as follows:

The 3-point gap with Claude should be viewed objectively: every score on this leaderboard carries an error bar of ±9 to ±13, and 3 points is statistically on the same level, so it's more accurate to say "tied in the first tier." But the 17-point lead over Kimi K3 and the 22-point lead over its own old version both exceed the error range, making them genuine improvements.

The category scores are also very consistent: it ranks 1st in the data analysis and consumer products categories, 2nd in brand marketing, gaming, and simulation, and 3rd in content creation tools and reference design.

The cheapest at the top, the highest-scoring among the cheap

The real highlight of this leaderboard is the price. The leaderboard also lists the USD pricing of each vendor's API:

Qwen's output price is less than a quarter of Claude's, and only forty percent of Kimi's. The leaderboard's official team did a cost-performance analysis: calculated by blended invocation cost, Qwen sits at the position with the highest score. The old rule of code leaderboards was that the stronger the model, the more expensive; this time the one at the very front is also the cheapest.

For domestic users calling through Alibaba Cloud Bailian, the RMB price is 12 yuan for input and 36 yuan for output per million tokens, with cache hits as low as 1 yuan.

Not just building web pages: long tasks that run for five straight days

The 0902 version focuses on "coding plus collaboration" reinforcement training, and the official team released several long-horizon task cases. In a financial simulation, it learned while doing across more than 2,000 rounds of trading iterations, ultimately achieving 416,000 yuan in simulated returns, a 4.16x return, 38% higher than the second-place GLM 5.2 and a 152% improvement over its own previous generation Qwen3.7-Max.

The scientific research reproduction experiment is even more striking: given a machine learning paper and a batch of GPUs, it worked autonomously and continuously for about 125 hours, writing 7,600 lines of code, executing more than 1,100 steps of operations, and completing 33 rounds of GPU training. It first reproduced the paper's six main conclusions from scratch, then tried 18 improvement directions on its own, and the new method it ultimately found scored 2.7 points higher than the original paper on the AIME24 math benchmark. In a chip design test, it went through about 500 rounds of interaction, optimizing an encryption circuit from 8,298 logic gates all the way down to 678. On PaperBench, which tests paper reproduction ability, it scored 93.0, the highest among the models officially compared.

Topping the web leaderboard, but coding isn't all-around yet

One thing needs to be made clear: WebDev tests web development, meaning tasks like front-end pages, interactions, and design reproduction. On SWE-bench Pro, which is closer to maintaining large code repositories, Qwen3.8-Max scores 67.7, still behind Claude Fable 5's 80.0. It has topped the web-writing leaderboard, but there's still a gap when it comes to chewing through complex engineering repositories.

In addition, Qwen simultaneously open-sourced Qwen3.8-27B with 27 billion parameters, which scored 1595 on the same leaderboard for 9th place, under the Apache 2.0 license, so local deployment enthusiasts can reach the edge of the first tier just by paying for electricity.

This is the first time a domestic model has taken first place overall in an overseas code arena. What's even more painful for its rivals is that it also sells cheaper than anyone else.

Are you bullish on Qwen3.8 Max? Let's talk in the comments