Qwen4-Max: Alibaba's Biggest Model Yet Is Cheap, and Closed
Have you ever looked at the price of a top AI model and thought, this is a tool for companies, not for me? Frontier models keep getting smarter, and the bill keeps getting bigger. So when Alibaba unveiled its new flagsh
Have you ever looked at the price of a top AI model and thought, this is a tool for companies, not for me?
Frontier models keep getting smarter, and the bill keeps getting bigger. So when Alibaba unveiled its new flagship, Qwen4-Max, on Sunday, November 15, the number that caught my eye was not the size. It was the price.
Qwen4-Max went generally available on November 16, the first model of Alibaba's new Qwen4 family. Qwen built its name on open models developers could download and run themselves. This time the flagship is different: closed, big, and cheap.
A library where you only pull the books you need
The number sounds absurd until you understand the trick. Qwen4-Max has a 2.4-trillion-parameter backbone, but it only activates 71 billion parameters for each token it produces.
Think of it like a giant library. A normal model reads every book on every shelf for every question. A mixture-of-experts model walks in, finds the few shelves that matter, and only pulls those books down. The library is enormous, but the work for each question stays small.
It also reads more than text. With native vision it understands images and video, and it can hold a million tokens in one prompt, roughly a whole codebase.
Photo by Peter Herrmann on Unsplash
The price is the real story
Through QwenCloud's API, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens. For comparison, Anthropic's Claude Opus 5.5 runs $4 and $20, and Google's Gemini 4 Carbon costs $2 and $10, with its cheap price set to double on December 1.
Alibaba also says it uses up to about 50 percent fewer tokens than its predecessor, Qwen3.8-Max, for the same work. On its own Work360 evaluation, the average cost per task fell from $4.43 to $1.67, a 62 percent cut.
That last number deserves a note. Work360 is Alibaba's own benchmark, and nobody outside Alibaba has verified it yet. But the direction is clear: the race is no longer just about the smartest model. It is about who can run a capable model for the least money, and that matters when your coding agent makes hundreds of calls a day.
Safety was trained in, not bolted on
The other part that stood out was safety, because Alibaba talked about it as part of the training, not a filter added at the end.
According to Reuters, the training included security checks alongside coding tests, aimed at reducing code injection, memory-safety flaws, and broken access controls. Alibaba says it filtered the training data by purpose, keeping material for defensive work like finding vulnerabilities, while removing data meant to improve attack capabilities. Its internal review found no sign of capabilities posing large-scale cybersecurity risks.
On a test that checks whether generated code both works and is secure, Qwen4-Max's pass rate rose from 22 percent to 31.8 percent. That is still far from perfect, and it is Alibaba's own measurement, but it points at the right problem. A coding model that writes fast code with security holes is not helping anyone.
The catch
So what is the catch? The biggest one is the weights. Qwen4-Max itself is closed. You cannot download it and run it on your own machine, the way you could with earlier Qwen models. You use it through Alibaba's cloud, or its apps Qwen Studio, QwenWork, and Qoder. Alibaba says smaller Qwen4 models will be open weights later, but there is no date yet.
The second catch is trust. Every impressive number here comes from Alibaba itself: the benchmarks, the cost figures, the safety tests. No outside lab has verified them yet, and no external red team has tested the model. That does not make the claims false. It makes them early.
Still, this is a release worth watching. The open-weights champion just showed its next step is a giant model you rent by the token, priced well below its Western rivals, with safety built into the training. If independent tests back up even half of Alibaba's numbers, many developers will have a new default for the heavy lifting.
What would you build if the smartest model in the room stopped being the most expensive one?
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.