Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards Alibaba released Qwen4-Max, a 2.4-trillion-parameter mixture-of-experts model that activates about 71 billion parameters per token instead of about 95 billion, and said it cuts the average cost per task to $1.67 from $4.43 for Qwen3.8-Max, a 62% reduction. Alibaba attributed the savings to 35% fewer tokens in High effort reasoning mode and new list prices of $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6, and said training-time data filtering and code security checks raised the share of functions passing both functionality and security tests on Last Secure Code Bench from 22% to 31.8%. Greyhound Research chief analyst Sanchit Vir Gogia cautioned that the figure is a vendor benchmark, saying "Alibaba has published the mileage it got on its own road, and a buyer's bill depends on the road the buyer drives," while Security Brigade founder and CTO Yash Kadakia questioned whether offensive and defensive cybersecurity knowledge can be separated through training-data filtering. Alibaba has released Qwen4-Max, a 2.4-trillion-parameter flagship model, claiming it reduces the average cost of completing professional tasks by 62% compared with its predecessor while introducing additional safeguards against offensive cybersecurity capabilities and other high-risk uses. “According to Work360 — a proprietary benchmark evaluating performance on challenging industry tasks across 721 specialized fields — Qwen4-Max outperforms its predecessor, Qwen3.8-Max, overall while using 35% fewer tokens at High effort reasoning mode and reducing the average cost per task by around 60%,” Alibaba said https://www.alibabacloud.com/blog/alibaba-unveils-qwen4-max-new-ai-teammate-for-work-creativity-and-discovery 603783 in a blog post. The Qwen team’s announcement https://qwen.ai/blog?id=qwen4-max puts the average task cost at $1.67 for Qwen4-Max, compared with $4.43 for Qwen3.8-Max, representing a 62% reduction. The model keeps the 2.4-trillion-parameter mixture-of-experts backbone of its predecessor, Qwen3.8-Max, but activates about 71 billion parameters per token instead of about 95 billion, the company said. It supports a 1-million-token context window. The company attributed the savings to lower token use and new list prices of $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6. Anthropic lists Claude Opus 5.5 at $20 per million output tokens. Sanchit Vir Gogia, chief analyst at Greyhound Research, said enterprises should treat the figure as a vendor benchmark rather than an assured reduction in their own AI spending. The relevant measure, he said, is the cost of an accepted result, including retries, reasoning tokens, and employee time spent checking outputs. “Alibaba has published the mileage it got on its own road, and a buyer’s bill depends on the road the buyer drives,” Gogia said. Yash Kadakia, founder and CTO of Security Brigade, said enterprises should measure successful task completion rather than token prices alone. “An enterprise should benchmark what it costs to complete 1,000 real business tasks successfully, not what it costs to generate a million tokens,” Kadakia said. Alibaba also cited improvements across coding and agentic evaluations, including Code Arena and Agents’ Last Exam. In an automotive manufacturing demonstration, the company said Qwen4-Max reduced manual workpiece programming time from approximately 475 minutes to 24 minutes. Alongside the cost figures, Alibaba set out changes to what the model learns about hacking at the training-data stage. “For cybersecurity, we filter data according to the specific purpose of each task, retaining data for defensive tasks such as vulnerability verification, detection, and remediation, while removing data for offensive tasks aimed at enhancing attack capabilities,” the Alibaba blog post said. The company 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%. Its internal security experts found the model’s defensive capabilities had “improved substantially,” the post added. Kadakia questioned whether offensive and defensive cybersecurity knowledge can be reliably separated through training-data filtering. “One of the fundamental challenges with cybersecurity AI is that offensive and defensive knowledge are largely the same knowledge,” he said, noting that understanding SQL injection is necessary both to exploit and prevent it. Gogia said the distinction requires independent validation, recommending that CISOs seek assessments from government evaluators or established external red teams rather than rely solely on Alibaba’s internal testing. The announcement follows an October investigation by Canadian AI company Cohere, which reported https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model that an earlier Alibaba model, Qwen3.8-Flash-Next, attempted to send a test file to an Alibaba Cloud storage destination in Beijing during a security test. Cohere said its sandbox blocked the request. A subsequent investigation by Check Point https://research.checkpoint.com/2026/coheres-qwen-story-looks-more-like-model-jaggedness/ replayed the behaviour 64 times and found 52 different destination hosts. It suggested unpredictable model behaviour was a more likely explanation than a deliberately configured backdoor. The incident involved an experimental model, not Qwen4-Max. Besides, Government evaluators have begun testing Chinese models independently. In July, the US Center for AI Standards and Innovation found https://www.nist.gov/news-events/news/2026/07/caisi-assessment-zais-glm-52 that the safeguards on Z.ai’s GLM-5.2 allowed assistance with agentic cyber exploit development. Gogia said CISOs should treat Alibaba’s filter “as a promise until a government evaluator or an established outside red team has tested it.” For enterprises outside China, Gogia said the geographic scope of Alibaba’s global service, prompt-retention policies, and jurisdiction governing stored data should be considered before deploying the model for sensitive workloads. He pointed to prompt retention of up to 30 days and longer retention of flagged content, including under zero-retention arrangements. Kadakia said enterprises should examine where prompts, retrieved documents, tool outputs, and logs are processed, who can access them, and what permissions agents have across internal systems. “The biggest risk with enterprise AI is no longer simply what information employees type into a chatbot. It is what an autonomous agent can access, what decisions it can make, and what actions it can execute across connected enterprise systems,” he said. Alibaba also plans to release smaller Qwen4 models with open weights. Kadakia said developers modifying those models through fine-tuning should not assume the original security safeguards will remain intact. “Safety training is not a permanent property of a model,” he said, recommending fresh adversarial testing and controls over tool access for customised versions. Gogia similarly recommended reassessing each modified model rather than automatically inheriting the original model’s security assurances.