New tool3 min read
Qwen4-Max is official and Alibaba prices it per task

Alibaba officially launched Qwen4-Max on November 16, 2026, and its launch blog claims a 62% lower cost per task than the previous Qwen3.8-Max: $1.67 on average instead of $4.43. The per-token price in the Frankfurt region drops only about 14%; according to Alibaba, the rest comes from the model using far fewer tokens.
Key facts
- Alibaba officially launched Qwen4-Max on November 16, 2026: 2.4 trillion parameters, 71 billion active per token, a 1 million token context window.
- According to Alibaba, cost per task on its own Work360 benchmark falls 62%, from $4.43 to $1.67.
- Per token, Qwen4-Max is about 14% cheaper than Qwen3.8-Max in the Frankfurt region ($1.413/$4.24 instead of $1.65/$4.951 per million tokens).
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
Early this morning I wrote about Qwen4-Max when it only existed in Alibaba’s cloud docs: prices, a Frankfurt region, APIs that mirror OpenAI and Anthropic (Qwen4-Max lands in Frankfurt and the AI model becomes a spare part). Now the official launch blog is out. It contains one number I find more interesting than any benchmark.
What does Qwen4-Max cost per task?
According to Alibaba, the average cost per task on its in-house Work360 benchmark falls from $4.43 to $1.67. That’s 62% less. The price list alone doesn’t explain it: in Frankfurt, Qwen4-Max costs $1.413 per million input tokens and $4.24 per million output tokens, versus $1.65 and $4.951 for its predecessor. About 14% cheaper per token.
Alibaba says the rest comes from consumption. The new model needs fewer tokens for the same work:
- Work360 (office and professional tasks): 24% fewer at “Xhigh” and 35% fewer at “High” reasoning effort
- Terminal-Bench 4.0 (coding tasks in a terminal): 48% and 49% fewer
- Score on Work360: 52.69 instead of 47.74
The headline specs from the blog: 2.4 trillion parameters, 71 billion active per token, a 1 million token context window, native vision.
Why price per token is the wrong number to compare
Most teams compare AI models by the price list: dollars per million tokens. That’s like comparing agencies by hourly rate without asking how many hours they need per brief. A model that is cheap per token but thinks three times as long ends up more expensive. Alibaba now markets Qwen4-Max on the number that actually matters: cost per finished task.
If you work in languages other than English, there’s a catch most people miss. German, for example, uses more tokens than English, as I wrote here: the German surcharge in AI, paid in tokens. The $1.67 per task applies to Alibaba’s test tasks, not to your proposals, briefs and customer emails.
The catch: Alibaba built the test itself
Work360 is Alibaba’s own evaluation framework, 26 domains and 721 specialized fields, developed by the Qwen team. That’s not an accusation, every vendor measures this way. But it means the 62% is a vendor claim on home turf. Whether it holds for your tasks, nobody knows, Alibaba included.
My rule of thumb for in-house versus outsourced fits right here. You outsource what you can brief, measure and swap. The model is such a part. What stays in-house is the test set you measure with. Without one, all you can do is trust vendor numbers, and that’s a poor basis for a budget.
What I would do now
- Collect 20 real tasks your team runs through AI every week, in your working language, with one example of a good result each.
- Measure per task: tokens in, tokens out, price, and whether the output is usable without rework. Rework is the most expensive line item.
- Compare cost per usable result, not dollars per million tokens. Only then is switching worth it, whether to Qwen, Claude or GPT.
How to set up proper metrics for AI work is covered in the guide AI marketing analytics; what agents can do in marketing today in AI agents for marketing. And why the model is becoming a spare part while the big vendors build bundles: the US will soon sell bundles, not models.
Got AI news? Submit news →
FAQ
How much does Qwen4-Max cost?
In the Frankfurt region, $1.413 per million input tokens and $4.24 per million output tokens (Alibaba pricing page, November 16, 2026). Per task, Alibaba cites an average of $1.67 on its own Work360 benchmark.
Is Qwen4-Max cheaper than Qwen3.8-Max?
Yes, according to Alibaba. About 14% per token in Frankfurt and 62% per task, because the model uses fewer tokens. The 62% is a vendor claim on a benchmark Alibaba built itself.
What is Work360?
An evaluation framework built by Alibaba for office and professional tasks, covering 26 domains and 721 specialized fields. Qwen4-Max scores 52.69 there, its predecessor 47.74.
Source
Qwen (Alibaba): Qwen4-Max: Your New Teammate Has Arrived
More sources
- Alibaba Cloud Model Studio – Preise je 1 Mio. Tokens (Stand 16.11.2026, im Artikel vom 16.11. 04:40 geprüft): qwen4-max Frankfurt (Global) 1,413/4,24 $, qwen3.8-max Frankfurt 1,65/4,951 $ → rund 14 % günstiger
- I4U KI-News – Vorgängerartikel 16.11.2026 04:40 (Doku, Preise, Schnittstellen)



