The battle of AI models is increasingly coming down to price. Alibaba made available the November 16, 2026 to all developers Qwen4-Maxthe new flagship model of its Qwen family, five days after a limited preview. The Chinese group puts forward a simple argument: doing the same work as the previous generation with much less calculation, and therefore for much less cost. Promises to be examined closely, because they are based on internal tests, and access conditions which deserve the attention of European companies.
What is Qwen4-Max
Qwen4-Max is based on a new architecture, called Qwen4. It is a type model expert mix: it counts 2,400 billion parametersbut only activates about 71 billion for each token processed. According to SiliconANGLE, it also relies on 328 billion so-called “N-gram” parameters, which play the role of a memory to avoid making repetitive calculations.
Concretely, an expert mixture model is divided into specialized subnetworks. For each piece of text (a token, or approximately three-quarters of a word), only a few experts are called upon. The model can thus be very large while remaining fast and less expensive to run.
The model reads the text, images and video and responds in text. He accepts until a million tokens in one go, enough to analyze an entire code base or a stack of long reports, without having to cut them into pieces. Alibaba intends it for programming, office work, web interface design and agent taskswhere AI connects many steps with tools. It also handles web search, function calling and structured responses.
The bet on efficiency
Alibaba’s central argument is not a record score, but calculation economics. According to the group, Qwen4-Max uses 24 to 49% fewer chips than its predecessor Qwen3.8-Max according to testing and reasoning tuning. On Work360an in-house test covering tasks from 721 professions, the average cost per task would pass from 4.43 to 1.67 dollarsa drop of around 60%.
This figure matters because a model charges for every token it writes, including those in its intermediate reasoning. A more concise model therefore costs less to use, even at the same price.
| Characteristic | Qwen4-Max |
|---|---|
| General availability | November 16, 2026 |
| Settings | 2,400 billion, including approximately 71 billion assets per token |
| Context window | 1 million tokens |
| Entries accepted | Text, images, video |
| Price (Singapore region) | $1.50 per million tokens in, $4.50 out |
| Price (Frankfurt and other regions) | $1.413 in, $4.24 out |
| Model weight | Closed, access only in the cloud |
Performances to put into perspective
Several precautions are necessary. The Work360 test was designed by Alibaba and its method has not been published; the media also cite cost reductions ranging from 58 to 62% depending on the version. Above all, as the specialized site DataNorth points out, the published comparisons relate to Qwen3.8-Maxand not on competing models of GPT, Claude Or Gemini.
Alibaba claims sixth place on Code Arenaa public ranking based on blind comparisons in interface programming, and a presence in the top group of two other rankings, without the corresponding scores having been detailed. The model is therefore among the good models of the moment, without claiming to dominate the best American models, such as Claude Fable 5.2 and Mythos 5.2.
The real advantage is price. Digital Trends points out that Claude Opus 5.5 is billed at $4 per million tokens for input and $20 for output: at this rate, the release of Qwen4-Max costs less than a quarter. For intensive uses such as programming agents, the difference can weigh heavily on the bill.
Security: declared safeguards, not yet verified
On cybersecurity, Alibaba says it has sorted the training data according to their purpose: it retained content useful for defense, such as finding and correcting vulnerabilities, and removed those used to improve attack capabilities. Based on its internal testing, the model does not have any capabilities that pose a large-scale cybersecurity risk.
The group also claims that Qwen4-Max writes safer code: on a code test that is both functional and secure, its success rate would increase from 22 to 31.8%. Specific training would have been carried out on chemical, biological, radiological and nuclear risks.
Using Qwen4-Max from Europe: the point of vigilance
Qwen4-Max is accessible through the programming interface of QwenCloud andAlibaba Cloud Model Studioas well as in the Qwen Work, Qoder and Qwen Studio tools. The price is the same with or without reflection mode, but this reasoning, enabled by default, is billed like exit tokens.
For a European company, the region of Frankfurt seems the natural choice. But DataNorth notes an important subtlety: Alibaba distinguishes between the location of data storage and the location where the model performs its calculations. In Frankfurt, Qwen4-Max is only offered with a global deployment scopewhich does not guarantee that the calculations remain in the European Union. The option that maintains treatments in the region exists, but is not offered for this model.
Practical consequence: for personal data or customer documents, the use of Qwen4-Max involves processing outside the Union, subject to the GDPR rules on data transfers. A caution that joins other recent limits aimed at Europe, such as Baidu’s ERNIE 5.5 license which excludes the European Union.
A Chinese offensive on prices
Qwen4-Max is part of a clear strategy of Chinese laboratories: to offer models close to the best for a fraction of the price, and to occupy the territory of cost-sensitive companies. Alibaba is also playing the openness card, promising Smaller Qwen4 models with open weightswithout a precise timetable, and by publishing in open source code an extension which allows the model to deliver a first result quickly on long tasks, then to improve it in stages.
For users, this competition is good news: it pushes the entire market to lower its prices. For European companies, however, it imposes a trade-off between cost, verified performance and control of the location of data processing. The coming months will tell whether independent tests confirm Alibaba’s promises, and whether the group opens a strictly European processing option.

