Qwen 4 Max, Tested: Alibaba Promised 60% Cheaper. On Six Real Builds, It Was 57%.
What Alibaba released
Alibaba has opened Qwen 4 Max to everyone, the flagship and first model of the Qwen 4 family. After a limited preview that started on November 11, it is now available on Qwen Cloud through the API or a token plan, and inside Alibaba's own apps: Qwen Work, Qwen Studio and the coding tool Qoder.
It is a mixture-of-experts model with 2.4 trillion parameters, of which about 71 billion are working at any moment. It reads images and video, and its context window is one million tokens, roughly 1,500 printed pages in one prompt.
The price claim
The price is $1.50 per million tokens in and $4.50 out, down from $2 and $6 for Qwen 3.8 Max. Alibaba's headline promise is the same work for about 60% less money. On its own Work360 test, the average cost per task drops from $4.43 to $1.67, which is 62% less, because the model is cheaper per token and also writes fewer tokens: about a quarter fewer on its highest setting and a third fewer on high.
The showcases
Alibaba's demos are ambitious: 3D scenes built in Blender, games in Unity, Unreal and Godot, production lines and car parts planned with two car-industry partners, a short-film pipeline with WonderClip, an 86-page due diligence report from 198 messy files, and treatment steps proposed from a full intensive-care record. In science, it worked alone for 48 hours, screened 2.7 million candidate molecules and directed a robotic lab to produce a new coolant molecule. These are showcases picked by Alibaba, and Alibaba itself says a human still makes the call in medicine.
The most useful feature may be Live Delivery: instead of waiting hours for an agent to finish, you get a usable first version early, then improvements in steps, and you can stop it or change direction at any point. Alibaba is releasing that plugin as open source. The Max weights are not public, but smaller Qwen 4 models are coming as open source, with no date yet.
The independent numbers
Artificial Analysis puts Qwen 4 Max at 43 on its intelligence index, above the average of its class and three points behind Kimi K3.1 Pro. Claude Opus 5.5 and GPT-6.1 Sol score higher. Its cost per task there is $3.27, second cheapest in the chart after Sol and about two thirds less than Qwen 3.8 Max. One warning: it wrote 127 million tokens to finish that test, against about 79 million for a typical model. It is a talker, and that matters, because price per token is only half the bill. The other half is how many tokens the model decides to write, including thinking you never see.
Six builds, one try each
So I ran my own bill. Qwen 4 Max and Qwen 3.8 Max got the same six prompts already used on Opus 5.5, Sol and three Gemini models, word for word, default thinking setting, no second chances: a life animation, a voxel castle, a short film, a 3D car showroom, a driving game and a music studio.
Qwen 4 Max ran six out of six, with one visual flaw: dark holes in the lower body of the car. Qwen 3.8 Max ran two out of six. Its castle never passed the loading screen because of one syntax error, its film was mostly black from a shader that didn't compile, its car never appeared, and its driving game stopped on the title screen.
The bill
For the six builds, Qwen 3.8 Max wrote 341,000 tokens and cost $2.05. Qwen 4 Max wrote 196,000 tokens and cost $0.89: 57% less, close to the promised 60, and done in about an hour instead of more than two. Across all models on the same prompts, Sol is still cheapest at $0.77, then Qwen 4 Max at $0.89, Gemini 4 Argon at $1.52, Hydrogen at $1.75, Opus 5.5 at $2.70 and Gemini 4 Carbon at $5.21.
Verdict
Qwen 4 Max is better, faster and cheaper than its predecessor at the same time, which rarely happens in one release. It is not the smartest model, and it is not the cheapest. But six working builds out of six for about a third of what Opus cost is high-level work. One sample per build, so treat it as a comparison you can click through, not a benchmark. Every build is on the site.