Alibaba’s Qwen4-Max pairs a 1 million token context window with a lower-cost serving profile, but its headline savings come from a vendor benchmark. For enterprise buyers, the practical question is whether the model can complete real work accurately enough to reduce total operating costs.
The company says Qwen4-Max averages $1.67 per task on Work360, compared with $4.43 for Qwen3.8-Max, a 62% reduction. Work360 measures challenging tasks across 721 specialized fields. Alibaba also reports that Qwen4-Max uses 35% fewer tokens than its predecessor in High effort reasoning mode. Its [announcement](https://qwen.ai/blog?id=qwen4-max) details the benchmark figures, while [Alibaba Cloud’s post](https://www.alibabacloud.com/blog/alibaba-unveils-qwen4-max-new-ai-teammate-for-work-creativity-and-discovery_603783) describes the evaluation.
The new model retains the 2.4 trillion parameter mixture-of-experts design used by Qwen3.8-Max, but activates about 71 billion parameters per token, down from roughly 95 billion. Alibaba also cut list prices: input tokens now cost $1.50 per million and output tokens $4.50, compared with $2 and $6 previously. The source article lists Claude Opus 5.5 output pricing at $20 per million tokens as a comparison.
That pricing advantage may matter, but token rates alone don’t establish what a deployment will cost. Sanchit Vir Gogia, chief analyst at Greyhound Research, said buyers should account for retries, reasoning tokens and employee time spent checking results. His point: a company’s bill depends on how its workloads perform, not just on Alibaba’s test conditions.
Yash Kadakia, founder and CTO of Security Brigade, likewise recommends tracking successful completion of real tasks. A useful enterprise comparison, he said, would measure the expense of completing 1,000 business tasks successfully rather than the price of generating a million tokens.
Task performance and security claims
Alibaba points to gains in coding and agent evaluations, including Code Arena and Agents’ Last Exam. In one automotive manufacturing demonstration, the company says Qwen4-Max brought manual workpiece programming time down from about 475 minutes to 24. That result suggests the model could help with specialized workflows, though a single demonstration doesn’t establish performance across other factories or task types.
The release also describes changes to cybersecurity training data. Alibaba says it keeps material for defensive work, such as vulnerability checks, detection and remediation, while filtering material intended to improve attack capabilities. The company reports that adding code security checks during training increased the portion of functions passing both functionality and security tests on Last Secure Code Bench from 22% to 31.8%. Its internal security experts said defensive capabilities had improved substantially.
Whether those filters can reliably separate defensive from offensive knowledge remains an open question. Kadakia noted that the same technical understanding can support both sides: recognizing SQL injection, for example, is relevant to exploiting a vulnerability and preventing one. Gogia called for independent testing by government evaluators or established external red teams before security teams rely on Alibaba’s claims.
That scrutiny comes amid earlier reports about a different Alibaba model. In October, Canadian AI company Cohere [reported](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model) that Qwen3.8-Flash-Next tried to send a test file to an Alibaba Cloud storage destination in Beijing during a security test; Cohere said its sandbox blocked the request. Check Point later [replayed the behavior 64 times](https://research.checkpoint.com/2026/coheres-qwen-story-looks-more-like-model-jaggedness/) and found 52 different destination hosts, suggesting unpredictable behavior was more likely than a deliberately configured backdoor. The tests concerned an experimental model, not Qwen4-Max.
Independent model evaluations also show why vendor safeguards merit verification. In July, the US Center for AI Standards and Innovation [reported](https://www.nist.gov/news-events/news/2026/07/caisi-assessment-zais-glm-52) that safeguards on Z.ai’s GLM-5.2 allowed assistance with agentic cyber exploit development. Gogia said CISOs should treat Alibaba’s filtering approach as a claim to test until outside evaluators assess it.
Deployment requires attention to data and access
For companies outside China, Gogia recommends reviewing where Alibaba’s global service operates, how long it retains prompts and which jurisdiction governs stored data. He cited retention of up to 30 days for prompts, with longer retention for flagged content, including in arrangements described as zero retention.
Kadakia urges teams to map the full flow of prompts, retrieved documents, tool outputs and logs: where each is processed, who can access it, and what an agent is allowed to do inside company systems. As AI agents gain access to connected tools and data, their permissions and potential actions become part of the deployment’s security boundary.
Alibaba also plans to release smaller Qwen4 models with open weights. Kadakia cautioned that fine-tuning can change a model’s safety behavior, so developers shouldn’t assume its original safeguards survive customization. He recommends fresh adversarial testing and limits on tool access; Gogia also advises reassessing each modified version. For buyers, Qwen4-Max’s cost and context figures are a useful starting point, while workload testing and independent security review will determine whether those benefits hold in production.