Alibaba launched Qwen4-Max in China on Monday, its new flagship AI model, claiming it beats Anthropic’s Claude on coding tests while cutting the cost of each task by 62%.
The release lands in the middle of a fast run of new models from Chinese labs, and it turns on price as much as power. Alibaba is selling Qwen4-Max as a cheaper worker for long coding and office jobs, not just a smarter chatbot. At the same time, the company kept the flagship closed, a break with the open releases that made its Qwen family the most downloaded in the world.
Key Points
- Alibaba launched Qwen4-Max on Monday 16 November, the flagship of its new Qwen4 family.
- The company says the model beats Anthropic’s Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro on coding tests, but those results are its own and have not been checked by outside testers yet.
- Prices fall 25% from the last flagship: $1.50 per million input tokens and $4.50 per million output tokens.
- The model is closed. Smaller Qwen4 models will be released as open weight, Alibaba says, but it gave no date.
- Alibaba says it stripped offensive cyberattack skills from the model’s training data while keeping skills used for cyber defence.
What happened
Alibaba’s Qwen team put Qwen4-Max on general sale on Monday through its QwenCloud service, after a preview that started on 11 November for subscribers to its Token Plan. It is also in Qwen Studio, QwenWork and Qoder, and developers can call it through interfaces built to match OpenAI’s and Anthropic’s.
The model runs on a new design for the whole Qwen4 family. Alibaba says it has a 2.4-trillion-parameter backbone, but only 71 billion of those parameters switch on for each token it processes, 25% fewer than the model it replaces. Another 328 billion parameters sit in N-gram embeddings. It takes text, image and video as input, reasons by default, and holds a one-million-token context window.
The price is the sharpest part of the launch. Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens, a quarter less than Qwen3.8-Max. Cached input costs $0.075 per million tokens. Alibaba says the model also uses fewer tokens on the same work: about 48% to 49% fewer on the Terminal-Bench 4.0 coding test, and 24% to 35% fewer on its own Work360 test of professional tasks.
On that Work360 test, which Alibaba built itself across 26 fields of work and 721 specialist areas, Qwen4-Max scored 52.69 against 47.74 for Qwen3.8-Max. Alibaba puts the average cost of a task on the test at $1.67, down from $4.43, a fall of 62%. Its pass rate for writing secure code rose from 22.0% to 31.8% at function level.
The company’s own charts tell a mixed story. On a legal-work test, Qwen4-Max scored 21.3, ahead of Anthropic’s Opus 5.5 at 9.2 and OpenAI’s GPT-6 Astra at 7.1. On Terminal-Bench 4.0, it scored 48.5, behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2. It debuted at No. 6 on LMArena’s Code Arena WebDev board with a preliminary 1,750 points, the highest Chinese model there.
Alibaba showed the model building 3D scenes in Blender, making games in Unity and Unreal Engine, and producing financial research with traceable sources. In one company demo, programming a single machined part fell from about 475 minutes to 24. In a 48-hour run, the model screened 2.7 million candidate molecules and steered an automated wet lab through four rounds of design, synthesis and testing to produce a physical sample of a new coolant material for data-centre chips.
Why it matters
Cost is the pitch. Agent software lives or dies on the price of each step, and Alibaba is arguing that a model can be both stronger and cheaper to run. If the coding claims hold up once outside testers run their own checks, Chinese models move closer to the front rank of US labs, and the price of frontier AI falls for everyone building on it.
The timing is part of the story. Tencent’s Hunyuan 4 and Moonshot’s Kimi K3.1 Pro arrived in recent weeks, and Chinese firms are now releasing new models in quick succession. Bloomberg described the launch as China’s latest challenge to US leaders. That race also runs through hardware: forensic tests found Nvidia Blackwell chips inside six Chinese AI models despite US export bans, and Alibaba’s own chip unit, T-Head, has filed to list in Hong Kong.
The closed flagship is the tension inside the launch. Qwen built its name on open releases, and its family is the most downloaded open-weight line in the world. Now the best model in the family is hosted only, with smaller open versions promised but undated. Alibaba chairman Joe Tsai backed open-source AI at China’s Wuzhen Summit, and the company says the smaller Qwen4 models will follow that path. Developers who want the top model today must rent it, not download it.
Safety shaped the release too. Alibaba says it filtered the training data to remove offensive cyberattack skills while keeping skills used to find vulnerabilities, and added datasets and training aimed at chemical, biological, radiological and nuclear risks. Bloomberg reported upgraded security measures after a series of incidents that raised concerns over open-weight models, and noted similar moves from Tencent and Moonshot.
The launch fits a bigger build-out. Alibaba has made artificial general intelligence its primary goal, and chief executive Eddie Wu has set a target of 20 gigawatts of data-centre capacity around the world by 2032. The head of the Qwen team, Liu Dayiheng, laid out a road map in September covering Qwen 4, Qwen 4.5 and Qwen 5, with the next generation scaled to between five and ten trillion parameters.
What happens next
Outside testers will now run their own checks, and that is where the launch will be judged. No outside lab had published a full rerun of Alibaba’s benchmark numbers when the model went on sale, and the LMArena entry is still preliminary. If the coding scores survive that scrutiny, Qwen4-Max becomes a serious option for developers who now pay US prices.
The preview version of the model retires on 25 November, leaving the paid version as the only way in. The open question is when the smaller Qwen4 models arrive as open weight, and whether they carry the same coding strength at a size developers can run themselves. Alibaba’s next test comes sooner than that: the market will watch whether developers switch to the cheaper model, and whether its rivals answer with price cuts of their own.
Sources: Bloomberg (via The Edge), South China Morning Post, Dow Jones, Qwen blog, AlphaSignal, AI Letter. Time: Monday 16 November 2026.



