By using this site, you agree to the Privacy Policy and Terms of Use.
Accept
Logic & LayersLogic & Layers
  • Tools
  • Earn with AI
  • Productivity
  • Automation
  • Guides
Logic & LayersLogic & Layers
  • Privacy Policy
  • About
Search
  • Tools
  • Earn with AI
  • Productivity
  • Automation
  • Guides
  • About
  • Contact
  • Blog
  • Privacy Policy
  • Complaint
  • Advertise
© 2026 Logic and Layers. Ruby Design Company. All Rights Reserved.
Hands-on test of Alibaba Qwen4-Max on real user tasks
Tools

Qwen4-Max: the $4.43 AI task now costs $1.67

Editorial Team
Last updated: November 16, 2026 3:17 am
Editorial Team
Share
Qwen4-Max in testing: Alibaba's new flagship promises flagship output at a fraction of the token cost.

One AI task that cost $4.43 on Alibaba’s old flagship costs $1.67 on the new one. That’s the core claim behind Qwen4-Max, which went wide this weekend with a 1-million-token memory and a much smaller bill per job. Cheap is nice, but cheap only matters if the work still gets done right.

Contents
What Qwen4-Max actually isWhere the 62% actually comes fromThe independent numbersWhat cheaper per task looks like in real workWhere Qwen4-Max wins, and where it gets beatenHow to try Qwen4-Max todayShould you switch?

What Qwen4-Max actually is

Qwen4-Max is Alibaba’s biggest AI model yet, and the spec sheet reads like a rocket catalog. It has 2.4 trillion parameters in total. Here’s the part that actually matters: it only switches on about 71 billion of them for each token it processes, a design engineers call a mixture-of-experts model. Think of a huge company where a small specialist team handles your request, instead of dragging every employee into the meeting.

There’s also a lookup layer of 328 billion extra parameters called N-gram embeddings. Don’t worry about the name. It’s basically a shortcut that lets the model recall common word patterns without burning much computing power.

Three more things you get without configuring anything: native vision (it reads images and video, not just text), a 1-million-token context window (roughly 750,000 words held in memory at once), and “thinking” turned on by default. That last one means the model reasons through hard problems before answering, and you can dial the effort from low up to extra-high depending on how much brainpower a task deserves.

Where the 62% actually comes from

The 62% isn’t a coupon code. It’s two savings stacked: the model uses fewer tokens per job, and each token got cheaper at the same time.

On Work360, Alibaba’s benchmark of professional work across 721 job specialties, Qwen4-Max finished tasks with about 24% fewer tokens than Qwen3.8-Max at the highest reasoning effort, and roughly 35% fewer at the standard high setting. On Terminal-Bench 4.0, a public test of agent work in a terminal, token use dropped by almost half.

Then the rates themselves came down:

Token type Qwen3.8-Max Qwen4-Max Change
Input (per 1M tokens) $2.00 $1.50 25% cheaper
Output (per 1M tokens) $6.00 $4.50 25% cheaper
Cached input read $0.25 $0.075 70% cheaper

Stack those effects and you get the headline: average cost per task on Work360 fell from $4.43 to $1.67. That cached-input line is the sneaky-good detail for anyone building on top of this, because long system prompts and repeated instructions stop being expensive once they’re cached. It’s the same logic behind decision models, which route easy work to cheap models so you stop overpaying for simple tasks.

The independent numbers

Alibaba graded its own homework, so here’s the outside check. Artificial Analysis scored Qwen4-Max at 43 on its Intelligence Index, up from 34 for the old model, and measured its cost per task falling from $10.02 to $3.27. Vals.ai ranked it the highest-scoring Chinese model on its index (#11 of 38 overall). So the direction holds: smarter and cheaper than its predecessor, Qwen3.8-Max. Still not the absolute best model on the market, and Alibaba’s own comparison charts openly show it losing several tests to Opus 5.5 and Kimi K3.1.

What cheaper per task looks like in real work

Benchmark charts are one thing. Alibaba also showed the model doing actual jobs, and the pattern across them is the same: less babysitting, fewer burned tokens.

The strangest one is also the most interesting. The team ran the model continuously for 48 hours on a chemistry problem, building a candidate pool of 2.7 million molecules and cycling through four rounds of design, test, and refine. It ended with a coolant molecule that passed real experimental validation. No human sat there approving each step. Whether you care about chemistry or not, that’s the shape of long agent work getting cheaper: the bill per attempt dropped, so running more attempts became affordable.

A more relatable example comes from manufacturing. Programming one machine workpiece used to take about 475 minutes of skilled manual effort. With Qwen4-Max handling the task, that fell to 24 minutes. If you sell services priced by the hour, notice what that does to your math. Deliverables get cheaper to produce, and the advantage shifts to whoever spots the change first.

There’s a workflow detail hiding in the launch notes too, called Live Delivery. Instead of making you wait for one giant perfect answer, the model ships a rough usable version first, then improves it in rounds based on your feedback. Early result in minutes, polished result after. For anyone who has watched a progress bar for ten minutes only to say “no, not that,” this is a genuinely better way to work with an AI.

Where Qwen4-Max wins, and where it gets beaten

Honest table time. These figures come from Alibaba’s launch charts plus third-party trackers:

Benchmark What it measures Qwen4-Max The competition
DeepSWE v1.1 Software engineering agents 75.7 Kimi K3.1: 75.5, GPT-6.1 Sol: 75.2, Opus 5.5: 74.2
Agents’ Last Exam Long multi-step agent work 46.7 Opus 5.5 and Kimi K3.1: 43.4
HLAB-Vals-1.1 Legal agent tasks 21.3 Kimi K3.1: 11.7, Opus 5.5: 9.2
Terminal-Bench 4.0 Terminal agent work 48.5 Kimi K3.1: 67.2, Opus 5.5: 59.6 (Qwen trails here)

Two honest notes worth your attention. First, on a security-focused test called Secure Code Bench, the share of passing code jumped from 22% to 31.8%, which is real progress, yet it also means most generated code still fails some checks. Second, early hands-on reviews of the preview build found the model’s visual taste (webpage aesthetics, rendered scenes) trails Anthropic’s Fable 5.2. Coding, yes. Art direction, not so much.

How to try Qwen4-Max today

You don’t need a budget approval to kick the tires. Free route first, then the paid paths:

  1. Chat for free. Open Qwen Studio, Qwen’s chat interface, pick Qwen4-Max in the model selector, and throw a real task at it. This is the fastest way to judge whether it fits your work.
  2. Pay per token. The API lives on QwenCloud at $1.50 per million input tokens and $4.50 per million output tokens. Both OpenAI-compatible and Anthropic-compatible endpoints are available, so most existing scripts just need a base URL and key swap.
  3. Inside tools you may already use. The model ships inside Qoder (Alibaba’s coding tool), QwenWork, and it’s listed on OpenRouter if you route models through there.
  4. Set the effort dial. Thinking runs at high by default. Bump tricky tasks to extra-high, drop repetitive ones to low, and watch the token bill shrink.

One limitation to know upfront: the Max model’s weights are closed, so you can’t download Qwen4-Max and run it on your own machine. Alibaba says smaller Qwen4 models will get open-weight releases later, though no date yet. If local is your thing, its predecessor still runs well on a gaming PC for free.

Should you switch?

Short verdict: yes for high-volume agent work, hold off only where it measurably loses.

If your month is full of document analysis, research pipelines, or long agent sessions, test Qwen4-Max this week. The cost-per-task drop is the single biggest price move among flagship models this quarter, and independent trackers confirm the direction. If you live in a terminal running coding agents all day, Kimi K3.1 still posts stronger numbers there, so run your own tasks before moving anything. And if you’d rather not send data anywhere, a local setup on your own hardware remains the zero-cost-per-token answer.

For everyone else, keep it simple: try the free chat, give it your hardest real task, and compare against what you use now. One afternoon of testing beats ten blog posts, including this one, honestly.

Do this now: pick one task you pay an AI to do every week. Run it on Qwen4-Max and on your current model, then compare output quality and token cost. If quality holds while the bill drops by half or more, you just found free budget.

You Might Also Like

Gemini Omni: Google’s AI video maker (free route inside)
Nvidia RTX Spark AI PCs: what they are and if you need one
Which Claude model should you use? (2026 guide)
Runway WorldPrompt: build AI worlds by describing them
Firefox Smart Window gets Mistral: how to try it today
TAGGED:AI AgentsAI ModelsAlibabaAPI costsqwen
Share
Previous Article Illustration of a shopping cart surrounded by product data points and indexes Amazon World Product Graph: what it means for sellers
Next Article smart home hub device standing on a living room shelf next to decor UGREEN HomeAgent: local AI for your whole smart home
Leave a Comment

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

banner banner
Create an Amazing Newspaper
Discover thousands of options, easy to customize layouts, one-click to import demo and much more.
Learn More

Latest News

Gemini 4 Hydrogen key art from Google's launch announcement
Gemini 4 Hydrogen: Google’s new fast AI model explained
Tools
Coordinator chat directing parallel AI worker threads on separate branches
Claude Code Projects just learned to run AI workers for you
Tools
smart home hub device standing on a living room shelf next to decor
UGREEN HomeAgent: local AI for your whole smart home
Automation
Illustration of a shopping cart surrounded by product data points and indexes
Amazon World Product Graph: what it means for sellers
Earn with AI

Recent Posts

  • Gemini 4 Hydrogen: Google’s new fast AI model explained
  • Claude Code Projects just learned to run AI workers for you
  • UGREEN HomeAgent: local AI for your whole smart home
  • Qwen4-Max: the $4.43 AI task now costs $1.67
  • Amazon World Product Graph: what it means for sellers

Recent Comments

  1. Gemini 4 Hydrogen: Google's new fast AI model explained on Gemini 4 Argon vs GPT-6: which one deserves your money?
  2. Claude Code Projects just learned to run AI workers for you on Claude Code cross-session messaging: stop copy-pasting between sessions
  3. UGREEN HomeAgent: local AI for your whole smart home on Notion AI data breach: what leaked and what to do
  4. Qwen4-Max: Alibaba's AI teammate that costs 62% less on Ollaya decision models: run Jev-style triage locally
  5. Amazon World Product Graph: what it means for sellers on Google Universal Cart: how the AI shopping cart works

You Might also Like

Doctor reviewing CT scan images on a medical display
Tools

Damo Radar: Alibaba’s open-source AI that spots cancer on CT scans

Editorial Team
Editorial Team
10 Min Read
Two terminal windows running Claude Code sessions side by side on a laptop
Productivity

Claude Code cross-session messaging: stop copy-pasting between sessions

Editorial Team
Editorial Team
10 Min Read
Instacart logo, maker of the Clementine AI grocery assistant
Tools

Instacart Clementine: the AI shopping assistant explained

Editorial Team
Editorial Team
8 Min Read
//

We influence 20 million users and is the number one business and technology news network on the planet

Quick Link

  • PRIVACY NOTICE
  • YOUR PRIVACY RIGHTS
  • INTEREST-BASE ADSNew
  • TERMS OF USE
  • OUR SITE MAP

Support

  • ADVERTISE
  • ONLINE BESTHot
  • CUSTOMER
  • SERVICES
  • SUBSCRIBE

Categories

  • Tools
© 2026 Logic and Layers. All Rights Reserved.