AboutAI ToolsCursorClaude CodeLovableBolt.newAI tool reviewsTutorialsVibe codingAI agentsMobile apps with AIAll tutorialsBlogSponsorWork With MeJoin 285K+ Subscribers

Video tutorial · The Metaverse Guy

I Tested Qwen 4 Max: Alibaba’s 2.4T AI Model vs Claude (10x Cheaper?)

Alibaba's latest model claims to beat Claude Opus 5.5 and GPT-6.1 Sol at agentic coding. I tested it against Claude Fable 5.2 with real-world projects and compared the cost.

I Tested Qwen 4 Max: Alibaba’s 2.4T AI Model vs Claude (10x Cheaper?)
Published
16 November 2026
Length
13 min
Views
20,968 views
On YouTube
I Tested Qwen 4 Max: Alibaba’s 2.4T AI Model vs Claude (10x Cheaper?)

Qwen 4 Max is a sparse mixture-of-experts model with 2.4 trillion total parameters and 71 billion active per token. I put it through a series of coding challenges in Kilo Code, from building a racing game from a screenshot to a 3D Rubik's Cube simulator, and compared each with Claude Fable 5.2. Here's how it performed and whether the 10x cost saving is worth it.

Qwen 4 Max: Specs and First Impressions

Alibaba just released Qwen 4 Max, its most capable AI model yet. It is a sparse mixture-of-experts model with a staggering 2.4 trillion total parameters, but only 71 billion are active per token, making it efficient despite its size. The context window is 1 million tokens, and it can output up to 262,000 tokens in one go. It accepts text and image inputs natively, and you can adjust its reasoning effort to balance speed and depth. The weights are closed, though Alibaba says smaller Qwen 4 variants will be open source later.

Benchmark Reality Check

Alibaba's launch numbers are impressive in places: DeepSWE v1.1 gives it 75.7, beating Claude Opus 5.5 (74.2) and GPT-6.1 Sol (75.2). On Agents’ Last Exam v1.1 it scores 46.7 versus Opus 5.5’s 43.4. But it stumbles on Terminal-Bench 4.0 (48.5 vs Kimi K3.1’s 67.2) and ProgramBench (31.5 vs Opus 5.5’s 65.0). Overall, it wins only 5 of the 16 benchmarks Alibaba shared. The independent Vals AI Index places it at #11 out of 38 — the top-ranked Chinese model, but not a clean sweep.

A Word from Our Sponsor

A quick thanks to fal for sponsoring this video. If you’re building AI agents, check out fal Agent using the link in the description.

My Coding Test Setup

I ran all my tests inside VS Code with the Kilo Code extension. This let me use both Qwen 4 Max and Claude Fable 5.2 with the exact same prompts, and I tracked the cost of every task.

Building Real Applications

First, I gave Qwen 4 Max a screenshot of a simple racing game and asked it to recreate the game. It produced a working canvas-based racer with moving obstacles and scoring. I then repeated the same prompt with Claude Fable 5.2 for a direct comparison.

Next, I requested a 3D environment using Three.js. Qwen 4 Max generated a navigable scene with lighting, a ground plane, and some basic objects. The task gave me a clear sense of the cost per task, which was dramatically lower than Claude’s.

For the interior design tool, the model had to build an interactive room planner where you could place and move furniture. Qwen 4 Max delivered functional drag-and-drop logic, though the UI was basic. The 3D Rubik’s Cube simulator followed — it rendered a cube you could rotate, though the twist animations sometimes lagged.

Finally, I asked for a 2D driving game with a top-down view. Qwen 4 Max built one with car physics and a scrolling road, but I also built the same game with Claude Fable 5.2 to compare the quality and polish side by side.

  1. Racing game from a screenshot
  2. Three.js 3D environment
  3. Interior design tool
  4. 3D Rubik's Cube simulator
  5. 2D driving game

Head-to-Head with Claude Fable 5.2

In both the racing game and the 2D driving game, Claude Fable 5.2 produced slightly more refined results — better collision detection, smoother handling, and nicer visuals. However, Qwen 4 Max was by no means unusable. It completed every task I threw at it, and the cost difference was stark: for projects where you iterate a lot, the savings add up quickly.

Final Verdict

Qwen 4 Max is a serious contender for agentic coding, especially if you are watching the budget. It excels on longer tasks where its large context window shines, and it wins some high-profile benchmarks. But it is not a straight replacement for Claude Opus 5.5 across the board — its losses on ProgramBench and Terminal-Bench matter. If you need the absolute best performance and are willing to pay, Claude still has an edge. If you want something 10x cheaper that can still build working apps, Qwen 4 Max deserves a spot in your toolbox.

Tools used in this video

  • fal Agent
  • Kilo Code
  • VS Code
  • Qwen 4 Max
  • Claude Fable 5.2
  • Three.js

Questions

What is Qwen 4 Max?
Qwen 4 Max is Alibaba's latest AI model, a sparse mixture-of-experts with 2.4 trillion total parameters and 71 billion active per token. It has a 1M token context window, native vision, and adjustable reasoning effort.
How does Qwen 4 Max compare to Claude Opus 5.5 in benchmarks?
It beats Claude Opus 5.5 on some agentic coding benchmarks like DeepSWE (75.7 vs 74.2) and Agents' Last Exam (46.7 vs 43.4), but falls behind on ProgramBench (31.5 vs 65.0) and Terminal-Bench.
Is Qwen 4Max open source?
The full Qwen 4 Max model is closed-weight, but Alibaba has promised that smaller Qwen 4 models will be released as open source.
How much does it cost to use Qwen4 Max?
Input costs $1.50 per million tokens and output costs $4.50 per million tokens, making it significantly cheaper than models like Claude.

Topics

Chapters

  1. 0:00 Qwen 4 Max is here
  2. 0:24 What is Qwen 4 Max?
  3. 0:34 2.4T parameters, 71B active
  4. 0:44 Specs: vision, 262K output, reasoning
  5. 0:53 Closed weights
  6. 1:04 Benchmarks: DeepSWE
  7. 1:24 Its real win: long tasks
  8. 1:37 Where it loses
  9. 1:57 5 of 16 benchmarks
  10. 2:04 Independent test: Vals AI
  11. 2:21 Sponsor: fal
  12. 4:41 My coding tests setup (Kilo Code + VS Code)
  13. 5:13 Test 1: Racing game from a screenshot
  14. 6:24 Qwen 4 Max vs Claude Fable 5.2
  15. 7:12 Test 2: Three.js 3D environment
  16. 7:45 Cost per task
  17. 8:09 Test 3: Interior design tool
  18. 9:46 Test 4: 3D Rubik’s Cube simulator
  19. 10:13 Test 5: 2D driving game
  20. 11:09 Fable 5.2 vs Qwen 4 Max: 2D game
  21. 11:57 Racing game comparison
  22. 12:47 Is Qwen 4 Max worth using?
  23. 13:11 Final verdict
Video description and links
Try fal Agent, try it here: https://fal.ai/agent?utm_campaign=fal_agent_creator_program&utm_source=youtube&utm_medium=influencer&utm_content=themetaverseguy_agent_integration_v1 Alibaba just released Qwen 4 Max, its most capable AI model yet: 2.4 trillion parameters, a 1 million token context window, and pricing of $1.50 / $4.50 per million tokens. Alibaba says it beats Claude Opus 5.5 and GPT-6.1 Sol at agentic coding, so I put it through my own coding tests to find out. In this video I break down what Qwen 4 Max actually is (a sparse mixture-of-experts model with only 71B parameters active per token), go through the benchmarks honestly (where it wins and where it loses), and then build real projects with it inside Kilo Code in VS Code: a racing game from a screenshot, a 3D Three.js environment, an interior design tool, a 3D Rubik’s Cube simulator and a 2D driving game. I also compare several of them against Claude Fable 5.2 using the exact same prompts, and show what each task cost. Qwen 4 Max key specs: 2.4T total parameters, 71B active per token (sparse MoE) 1M token context window, up to 262K output tokens Text + image input (native vision) Adjustable reasoning effort $1.50 input / $4.50 output per 1M tokens Closed weights (smaller Qwen 4 models promised as open source) Benchmarks mentioned (Alibaba’s launch numbers unless noted): DeepSWE v1.1: 75.7 (Kimi K3.1 75.5, GPT-6.1 Sol 75.2, Opus 5.5 74.2) Agents’ Last Exam v1.1: 46.7 (Opus 5.5 43.4) Terminal-Bench 4.0: 48.5 (Kimi K3.1 67.2) ProgramBench: 31.5 (Opus 5.5 65.0) Vals AI Index (independent): #11 of 38, the top Chinese model Questions? Leave them in the comments.

This article was written from the video's description with the help of AI. Watch the video for the full walkthrough.

Related tutorials