- Subject
- Microsoft 365 & SharePoint
- Filed
- Source
- James Smith
- Loans
- 750
Alibaba’s Qwen4-Max Cuts Benchmark Task Costs 62%, but Security Needs Testing
Alibaba reports lower task costs and added cyber safeguards for Qwen4-Max. Enterprise teams should validate both claims against their own workloads.

Alibaba says its Qwen4-Max flagship model can complete benchmarked professional tasks at an average cost 62% below its predecessor. For IT teams weighing that claim, the useful question is whether the savings hold on their own work, after retries, output checks and staff review are counted. The model also adds training safeguards intended to limit offensive cyber capabilities, a claim that calls for testing before teams rely on it in sensitive environments.
What Alibaba says changed
Qwen4-Max has 2.4 trillion parameters, the same mixture-of-experts backbone as Qwen3.8-Max. Alibaba says the newer model activates about 71 billion parameters per token, down from about 95 billion, and supports a context window of 1 million tokens.
The Qwen team puts average task cost at $1.67 for Qwen4-Max, compared with $4.43 for Qwen3.8-Max. The company attributes the difference to lower token use and reduced list prices: $1.50 per million input tokens and $4.50 per million output tokens, compared with $2 and $6 previously. Alibaba also says Qwen4-Max uses 35% fewer tokens in High effort reasoning mode.
The figures come from Work360, Alibaba’s proprietary benchmark of challenging tasks across 721 specialized fields. The company says Qwen4-Max outperforms Qwen3.8-Max across that benchmark. Those results describe Alibaba’s test conditions, though; they don’t establish what a particular organization will spend or how often its users will need to correct model output. Alibaba’s product description and the Qwen team’s announcement provide the company’s details.
That distinction matters when comparing model costs. Greyhound Research chief analyst Sanchit Vir Gogia argues that enterprises should measure the price of an accepted result, including retries, reasoning tokens and employee time spent reviewing responses. Security Brigade founder and CTO Yash Kadakia makes a similar point: a useful comparison is the cost of successfully completing 1,000 real business tasks, rather than the price of generating a million tokens.
Alibaba also reports gains on coding and agentic evaluations, including Code Arena and Agents’ Last Exam. In an automotive manufacturing demonstration, it says Qwen4-Max cut manual workpiece programming time from approximately 475 minutes to 24 minutes. That is a substantial result for the demonstrated task, but teams should check whether their own processes, inputs and acceptance criteria resemble the demonstration.
Cyber safeguards remain a claim to validate
Alibaba says it filtered training data by task purpose, retaining material for defensive work such as vulnerability verification, detection and remediation while removing material aimed at strengthening attack capabilities. The company also says it added code security checks during training. On Last Secure Code Bench, it reports that the share of functions passing both functionality and security tests rose from 22% to 31.8%.
Alibaba’s internal security experts described the model’s defensive capabilities as substantially improved. That assessment is useful context, but it is still the company’s own evaluation. Kadakia cautions that offensive and defensive cyber knowledge overlap: understanding SQL injection, for example, can help a practitioner find a weakness or prevent it. Filtering training data by stated purpose therefore doesn’t, by itself, show how the model will respond to every prompt or tool-enabled task.
Gogia recommends independent assessments from established external red teams before security teams treat the safeguards as dependable. For a deployment review, that means examining the model’s behavior against the organization’s own threat scenarios, permissions and connected tools. Teams should also distinguish results for Qwen4-Max from reports about earlier experimental systems.
One such earlier-model report came from Canadian AI company Cohere, which said a test involving Qwen3.8-Flash-Next produced an attempt to send a test file to an Alibaba Cloud storage destination in Beijing. Its sandbox blocked the request. Check Point later replayed the behavior 64 times and found 52 different destination hosts; its assessment suggested unpredictable model behavior was more likely than a deliberately configured backdoor. That incident concerned an experimental model, not Qwen4-Max, so it should not be treated as a test result for the new release.
Review data handling and customized versions
For organizations outside China, Gogia says deployment reviews should cover the geographic reach of Alibaba’s global service, prompt retention and the jurisdiction that applies to stored data. He points to retention of prompts for up to 30 days, with longer retention for flagged content, including in arrangements described as zero-retention. Teams handling sensitive workloads should confirm the terms that apply to their specific service and configuration before sending data.
Kadakia recommends mapping where prompts, retrieved documents, tool outputs and logs are processed, who can access them, and which internal systems an agent can reach. The practical exposure depends on the model’s permissions as well as its generated text: an agent connected to enterprise tools may be able to retrieve information or take actions within the access it has been granted.
Alibaba also plans to release smaller Qwen4 models with open weights. Kadakia warns that fine-tuning or otherwise modifying a model can change its behavior, so teams shouldn’t assume the original safeguards will remain effective. He recommends fresh adversarial testing and controls on tool access for customized versions. Gogia likewise advises reassessing each modified model on its own merits.
For now, Qwen4-Max gives administrators two claims to verify in a controlled evaluation: lower cost for accepted work and safer behavior on cyber-related tasks. Track the full effort needed to complete representative business tasks, and test the exact model configuration, data handling and permissions the organization plans to use.