Qwen4-Max Cuts AI Task Costs by 60%: Is an AI Price War Brewing?

- Alibaba says Qwen4-Max reduced average task costs by 60% compared with its predecessor, Qwen3.8-Max.
- The new model uses fewer tokens and lower API rates, while activating 25% fewer parameters per token.
- Qwen4-Max’s cost advantage raises questions about AI pricing competition.
Alibaba is making cost efficiency a central part of the Qwen4-Max launch. The company says the model reduced average task costs by 60% compared with Qwen3.8-Max in an internal evaluation, while its API rates are also lower.
The figures point out a question for OpenAI, Anthropic and Google: could cheaper AI agents pressure rivals to cut prices, or will those companies need to show that their more expensive models deliver better results?
Qwen4-Max’s Cost Savings Put Efficiency in Focus
QwenCloud lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens. The previous flagship, Qwen3.8-Max, cost $2 and $6, respectively. Reused input served from an automatic cache is priced at $0.075 per million tokens.
Alibaba’s Work360 evaluation provides another measure of the change. It puts Qwen4-Max’s average cost per task at $1.67, down from $4.43 for Qwen3.8-Max. The company also reports that Qwen4-Max used 35% fewer tokens at high reasoning effort and 24% fewer at extra-high effort.
On Terminal-Bench 4.0, the token use fell by 49% and 48% at those respective effort levels. Alibaba says the new model performed better than its predecessor across nearly all the capabilities it evaluated.
The two models have the same 2.4-trillion-parameter scale, but Qwen4-Max activates 71 billion parameters per token, 25% fewer than Qwen3.8-Max. Alibaba links that architectural change to lower inference costs. Together, the company’s figures show that the reported task savings reflect both reduced token use and lower API rates.
Do Rival Models Justify Their Higher Prices?
The comparisons show American labs at the top of Arena’s overall agent and web development tables, while Qwen4-Max leads in some Chinese rankings. They also describe American models as considerably more expensive. However, they do not include rival prices or task-cost results for equivalent jobs.
That limits what the figures can show about competition. Alibaba’s 62% reduction compares Qwen4-Max with Qwen3.8-Max, not with models from OpenAI, Anthropic, or Google. API rates alone also cannot show how much each system costs to complete the same task. Without those comparisons, the launch doesn’t show whether rivals need to cut prices or whether their higher costs produce better results.
Closed Weights Leave Developers With a Trade-Off
Qwen4-Max is closed-weight, so developers cannot download its parameters or run it privately on their own hardware. Alibaba says smaller Qwen4 models with open weights are planned, but it has not announced their sizes or release dates.
That plan may matter to developers already building on Qwen. An August report from Hugging Face counted 151,448 Qwen-based derivative models on its platform, about 4.7 times the number built specifically on Meta’s Llama family.
For now, Alibaba’s figures show lower task costs within its own model line. Whether those savings lead to broader price cuts remains an open question until comparable task-cost data is available for rival models.



