Alibaba made Qwen4-Max, its new flagship AI model, available to every developer on 16 November, five days after a limited preview. You now pay per token through Alibaba Cloud Model Studio and QwenCloud, at $1.50 per million input tokens and $4.50 per million output tokens. Alibaba says the model needs far fewer tokens than Qwen3.8-Max, cutting the average cost per task by about 60% in its own tests.
Alibaba says Qwen4-Max needs up to half the tokens of Qwen3.8-Max
Qwen4-Max reads text, images and video and answers in text. It can take in up to 1 million tokens at once, enough for a large codebase or a stack of long reports. Alibaba aims it at coding, office work, front-end design and agent tasks, where a model works through many steps with tools. It also supports web search, function calling and structured output in JSON.
Like its predecessor, it is a mixture-of-experts model: only a small part of the model is active for each token. According to Alibaba’s press release, as reported by Business Standard, that is about 71 billion of 2.4 trillion parameters.
The main claim is efficiency. Alibaba reports that Qwen4-Max uses 24 to 49% fewer tokens than Qwen3.8-Max, depending on the test and the reasoning setting. On Work360, its own test of tasks from 721 professional fields, Alibaba puts the drop in average cost per task at about 60%. Fewer tokens per answer is a direct saving, because you pay for every token the model writes, including its reasoning.
Treat that figure as an indication. Work360 was built by Alibaba, and the method has not been published. News reports also quote different versions of the result, from 58% to 62%.
Alibaba’s only outside ranking so far is sixth place in Code Arena
Alibaba’s documentation and the launch coverage contain no head-to-head scores against GPT, Claude or Gemini. The comparisons are with Qwen3.8-Max. Alibaba says Qwen4-Max ranks sixth in Code Arena, a public blind test for front-end coding. It also claims a place in the leading group on Agents’ Last Exam and the Vals Index. The reports we read give no scores for those last two.
The weights stay closed, so Qwen4-Max only runs as a cloud service. Alibaba says smaller Qwen4 models will follow as open weights in stages, without dates.
On safety, Alibaba says it removed training material aimed at offensive hacking but kept material for finding and fixing vulnerabilities. Its internal tests found no significant high-risk capabilities, an assessment that rests on its own testing and expert review.
Qwen4-Max costs $1.50 per million input tokens, a little less outside Singapore
Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens in Alibaba Cloud’s Singapore region. In Frankfurt, Virginia, Tokyo, Hong Kong and Beijing it costs slightly less: $1.413 and $4.24. The price is the same with or without thinking mode, but the model’s reasoning is billed as output. Thinking is on by default.
- Model ID: qwen4-max, usable through OpenAI- and Anthropic-compatible APIs
- Cached input: $0.075 per million tokens in Singapore, $0.071 in the other regions
- Free quota: 1 million tokens for 90 days, in Singapore only
- Also in: Alibaba’s Token Plan subscription and its QwenWork, Qoder and QwenStudio apps
- The preview model, qwen4-max-preview, is retired after 25 November (Beijing time)
In Frankfurt, Qwen4-Max runs in a mode that can process prompts in China
For European organisations, the Frankfurt region looks like the obvious choice. The details matter. Alibaba stores your data in the region you pick, but a separate setting, the deployment scope, decides where the model does its computing. In Frankfurt, Qwen4-Max is only offered with the Global scope, which Alibaba’s Frankfurt documentation describes as “any available node, including Chinese mainland and overseas”.
Frankfurt also has an EU scope that keeps the computing in the region, but Qwen4-Max is not offered with it. The Singapore option excludes mainland China, but it runs outside the EU as well. For personal data or client files, both routes mean processing outside the EU, so the GDPR rules for data transfers apply.
Move to the qwen4-max model ID now, then count tokens on your own tasks
If you tested the preview, change the model ID before it is retired. Then compare it on your own work. In the agent and automation projects DataNorth builds, the bill depends on how many steps and tokens a task takes. The price per token matters less. Alibaba’s choice to report cost per task is the right yardstick, but its 60% comes from its own test.
Teams that run coding or agent workflows on GPT or Claude can test Qwen4-Max now on public or synthetic data. Run the same tasks on both, and compare the cost and quality of each completed task. Organisations that need client data processed inside the EU can wait until Alibaba offers Qwen4-Max with the EU scope. Teams that want to run a model on their own servers will have to wait for the smaller open Qwen4 models.
For more information, visit the official announcement of Qwen4-Max on the Qwen blog.