Same prompts. Same task. Two AI models. One cost nearly seven times more, and the cheaper model produced the stronger result.
Just hours after Alibaba unveiled its largest AI model to date, the new 2.4-trillion-parameter Qwen 3.8 Max has outperformed Fable 5 in a head-to-head coding test that asked both AI models to build three self-contained 3D physics simulations. The result gets more interesting once price enters the picture: Qwen generated 46,600 tokens for about $0.28, compared with 38,700 tokens and roughly $1.93 for Fable 5.
That means Qwen generated more code at around one-seventh the cost.
The comparison, published by Atomic Chat, goes beyond benchmark scores on a leaderboard. Atomic shared the prompts used for the test and posted a side-by-side video showing what each model produced. Anyone with access to the models can inspect the prompt, watch the resulting applications run, and attempt the test independently.
That makes the experiment unusually easy to scrutinize.
Qwen 3.8 Max And Fable 5 Were Asked to Build the Same Three 3D Scenes
Atomic gave Qwen 3.8 Max and Fable 5 a demanding software-generation task: create three individual HTML files containing self-running 3D scenes with working physics.
The first was a marble machine. The second recreated an automobile assembly line. The third simulated a sawmill processing logs.
“We gave two models the same task. Build three self-contained 3D scenes, each one HTML file with real physics that runs itself.
Prompts:
A marble machine that lifts marbles up a wheel and drops them on a loop A car factory assembly line A sawmill cutting logs into planks,” Atomic said in a post on X.
These are harder tasks than producing a static webpage. Each model had to generate code that coordinated objects, movement, geometry, physical behavior, lighting, textures, animation, and user-visible state inside a functioning browser experience.
Atomic didn’t just publish its verdict. It shared a side-by-side video showing the resulting simulations, giving viewers a chance to inspect what each model actually built. The published prompts make the test repeatable, too. Developers with access to both models can run the same instructions and compare their own results, though individual generations may vary.
The resulting video gives viewers something benchmark tables rarely provide: direct visual evidence.
Qwen 3.8 Max beat Fable 5 at building 3D physics scenes for 7x cheaper!
We gave two models the same task. Build three self-contained 3D scenes, each one HTML file with real physics that runs itself.
Prompts: – A marble machine that lifts marbles up a wheel and drops them on a… https://t.co/LIXtOLbO9M pic.twitter.com/Z6W8GnHPar
— atomic.chat (@atomic_chat_hq) August 3, 2026
In the marble-machine test, Qwen’s version shows marbles visibly traveling up the lifting wheel before moving through the rest of the mechanism. Fable 5’s output appears to skip part of that physical sequence, with marbles appearing near the top rather than visibly completing the lift.
That may sound like a small difference, but it gets to the heart of the task. The prompt asked the model to construct a machine whose parts behaved coherently. A simulation that visually skips a mechanical step has missed part of the requested behavior.
Qwen’s output presented a more complete interpretation of the system.
The Factory Test Shows a Bigger Gap
The automobile factory comparison pushes the models into more complicated territory.
Qwen produced a scene with recognizable vehicles, robotic equipment, welding activity, and a more polished production sequence. The output shown in Atomic’s video contains greater visual detail, richer textures, stronger shadows, and more continuous movement across the scene.
Fable 5 generated a functioning prototype, but the result appears simpler and less refined in the side-by-side demonstration.
That distinction matters for AI coding tools.
Developers increasingly care about whether a model can take a broad description and turn it into usable software without dozens of rounds of correction. A model that understands how parts of a simulated system relate to one another can save substantial development time, especially for prototypes, interactive demos, browser games, training tools, and visualization software.
The sawmill test tells a similar story. Qwen’s version shows logs moving through the processing system and includes operational statistics tied to the simulation. Atomic pointed to less polished physical behavior in Fable 5’s version.
Taken together, the three scenes suggest Qwen handled the combination of coding, visual structure, and simulated physical behavior better in this particular run.
The $0.28 Number May Be the Bigger Story
The visual comparison grabs attention, but the economics could matter more.
Atomic reported that Qwen 3.8 Max generated about 46,600 tokens at a cost of roughly $0.28. Fable 5 generated about 38,700 tokens for approximately $1.93.
Qwen generated around 20% more tokens, yet the run cost about 85% less.
That is an unusual combination.
AI model comparisons often involve a familiar tradeoff. A cheaper model may perform well enough, but developers accept weaker output in exchange for lower inference costs. Atomic’s test points to a different possibility: lower-priced models may begin matching or beating far more expensive systems on practical software-generation tasks.
For startups running thousands or millions of model calls, that gap can become meaningful very quickly.
A sevenfold difference in generation cost does not matter much for someone running one prompt. Multiply that difference across large coding agents, automated testing systems, AI development platforms, or enterprise workloads, and the economics start changing.
The question stops being, “Which model is best?”
It becomes, “How much capability can I buy per dollar?”
Qwen’s showing in this test makes that question harder for higher-priced AI providers to ignore.
Alibaba Is Pushing Qwen Deeper Into Coding and Agentic Work
Qwen 3.8 Max is Alibaba’s latest flagship model and comes with 2.4 trillion parameters in a mixture-of-experts architecture, with roughly 95 billion parameters active during inference.
Alibaba has positioned the model heavily around coding, long-running agent tasks, and multimodal work.
Atomic’s test gives developers a more tangible example of what those claims can look like outside standard evaluation suites.
Building an interactive marble machine from a text prompt requires more than generating syntactically correct JavaScript. The model has to translate a description of a mechanical process into geometry, motion, object relationships, and executable code.
The car factory adds another layer. Robots have to move around vehicles. Production events need visual cues. Cars need recognizable structure. The environment needs enough spatial consistency to make the scene believable.
The sawmill requires another set of relationships between logs, machinery, movement, and processing state.
Those tasks make the demonstration useful precisely because failure is easy to see.
A benchmark score may tell developers that one model scored 74.3 and another scored 71.8. A marble suddenly appearing at the top of a machine tells you something instantly.
The Published Prompt Makes the Test More Interesting
Atomic’s decision to show the prompt and resulting video gives the comparison greater value than a simple claim of victory.
Readers do not have to accept Atomic’s conclusion at face value. They can inspect what each model built.
Developers with access to the models can go one step further and rerun the prompts themselves.
That does not mean every run will produce identical results. Generative models can return different outputs based on sampling settings, model configuration, and randomness. A second run could narrow the gap, widen it, or produce a different winner.
That limitation matters.
This is one head-to-head demonstration, not a comprehensive evaluation proving that Qwen 3.8 Max is better than Fable 5 across coding, reasoning, or every other task.
Still, reproducible prompts plus visible outputs give the test a level of transparency that many viral AI comparisons lack.
The claim is not simply “Qwen won.”
The evidence is sitting on the screen.
Open Weights Could Make AI Economics More Disruptive
Another part of the Qwen story could matter just as much as the benchmark result.
Alibaba plans to release open weights for Qwen 3.8 Max, opening the door for developers and infrastructure companies to deploy optimized or quantized versions outside Alibaba’s hosted API.
Atomic Chat, which focuses on local AI inference, says it plans day-zero support for the model and expects quantized versions to appear on Hugging Face.
Running a 2.4-trillion-parameter model locally is far from the casual laptop experience associated with small open models. Hardware requirements, quantization choices and deployment architecture will determine what “local” means in practice for a model of this size.
Yet open weights change the economic conversation.
Hosted API prices are one thing. Giving developers the ability to deploy, optimize and operate a model on infrastructure they control creates a different competitive dynamic.
That matters for companies concerned about cost, data residency, latency or dependence on a single API provider.
A Warning Shot for Expensive Frontier Models
Atomic’s demonstration does not settle the Qwen versus Fable debate. One video cannot do that.
It does show something worth paying attention to.
A model priced far below a premium competitor was given the same prompts, generated more tokens and produced 3D simulations that appear more complete and visually convincing in the published comparison.
That is the kind of result AI companies cannot dismiss as a leaderboard curiosity.
Developers increasingly have access to capable models from Alibaba, DeepSeek and other labs at prices that would have looked improbable a short time ago. As those models improve at coding and agent-driven software creation, premium model makers face pressure from two directions at once: capability and cost.
Atomic’s test captures that shift in three browser windows.
A marble machine. A car factory. A sawmill.
And a $0.28 bill.
Alibaba’s Qwen 3.8 Max Beats Fable 5 in 3D Physics Coding Test at 7X Lower Cost
