Alibaba has released Qwen4-Max, a 2.4-trillion-parameter flagship model that the company says reduces the average cost of completing professional tasks by 62% compared with its predecessor, Qwen3.8-Max. Alibaba cites its proprietary Work360 benchmark, which evaluates performance on challenging industry tasks across 721 specialized fields, and says the model outperforms its predecessor overall while using 35% fewer tokens at High effort reasoning mode. The average task cost is put at $1.67 for Qwen4-Max, compared with $4.43 for Qwen3.8-Max.
The company attributed the savings to lower token use and to new list prices of $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6. The model keeps the mixture-of-experts backbone of its predecessor but activates about 71 billion parameters per token instead of about 95 billion, and it supports a 1-million-token context window. Alibaba also cited gains across coding and agentic evaluations, including Code Arena and Agents' Last Exam, and said that in an automotive manufacturing demonstration Qwen4-Max cut manual workpiece programming time from approximately 475 minutes to 24 minutes.S1S2
Analysts cautioned that the figure is a vendor benchmark rather than a guaranteed reduction in enterprise AI spending. Sanchit Vir Gogia, chief analyst at Greyhound Research, said the relevant measure is the cost of an accepted result, including retries, reasoning tokens and the employee time spent checking outputs. Yash Kadakia, founder and CTO of Security Brigade, argued that enterprises should benchmark what it costs to complete 1,000 real business tasks successfully rather than the cost of generating a million tokens.S1S2
Alongside the cost claims, Alibaba described changes to the model's cybersecurity training. The company said it filters training data by task purpose, retaining material for defensive work such as vulnerability verification, detection and remediation while removing data aimed at enhancing attack capabilities. It said code security checks added during training lifted the share of functions passing both functionality and security tests on Last Secure Code Bench from 22% to 31.8%, and that its internal security experts found the model's defensive capabilities had improved, though these security claims have yet to be independently tested.S1S2