Claude Opus 5 Builds Full 3D Worlds and Sims From One Prompt
Claude Opus 5 turns single prompts into 3D worlds, physics sims, and games. Here's what the model actually produced in hands-on testing.
What can Claude Opus 5 actually build from a single prompt?
Claude Opus 5, Anthropic’s latest frontier model, can generate playable 3D environments, physics simulations, and interactive tools from a single prompt, with almost no manual coding cleanup required. In hands-on testing, one user got the model to produce a walkable 3D gallery, a Kerbal-style orbital launch simulator, a cloth physics demo, a predator-prey ecosystem, a drone flight sim, and a sneaker configurator, all generated directly by the model rather than pulled from existing templates or assets.
TL;DR
- One-shot 3D generation is now viable: Opus 5 produced a full walkable 3D art gallery with sound cues and interactive elements from a single prompt, with what the tester called “virtually zero work.”
- Physics simulations look convincing, including a cloth simulation that rips and drapes realistically and a wrecking ball demo with adjustable mass and a live destruction percentage.
- Emergent systems work too, shown by a predator-prey ecosystem simulation where spawning more foxes crashed the rabbit population and then the fox population in turn, mimicking real ecological dynamics.
- Original asset creation, not scraping, sets this apart. In the sneaker configurator demo, Opus 5 built the shoe’s visual elements itself using SVG rather than compositing existing images, something models have historically struggled with.
- Cost per task is dramatically lower than comparable frontier models. On agentic computer-use benchmarks, Opus 5’s mid-tier reasoning setting matched a rival model’s output at roughly half to a fifth of the price.
- Benchmark scores back up the demos: Opus 5 scored 30.2% on ARC-AGI-3 (novel problem solving), versus roughly 2% for Opus 4.8 and GPT-5.6, and hit 26% on a new “Automation Bench” for business workflow generation, up from 17.4% on the prior generation.
- It is not perfect: a double pendulum demo showed a minor visual glitch, and agentic coding benchmarks still show GPT-5.6 ahead of Opus 5 in that specific category.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does Opus 5 generate 3D worlds and simulations from text?
The model writes the underlying code (typically JavaScript, WebGL, or similar web-based rendering stacks) in response to a prompt, then that code runs directly in a browser to produce an interactive scene. This is not image generation or video generation. It is Opus 5 acting as a coding agent, writing a program that simulates physics, renders 3D geometry, and handles user input, all in one pass.
What makes the outputs notable isn’t just that they exist. Simple 3D demos and physics toys have been possible with AI-assisted coding for a while. The distinction here is fidelity and originality: the sneaker configurator example is worth dwelling on, because instead of importing a shoe model or texture from somewhere online, Opus 5 constructed the visual elements itself. That kind of from-scratch compositing is a harder problem than remixing existing assets, and it’s historically been a weak point for coding models.
The ecosystem simulation is another good marker of underlying quality. A predator-prey model isn’t just a visual trick, it requires the model to correctly encode feedback loops: rabbits eat grass, foxes eat rabbits, and population sizes need to respond dynamically to user actions like spawning more animals. When the tester added foxes, the rabbit population collapsed, and the fox population followed it down shortly after, which is the behavior you’d expect from a correctly modeled system, not a scripted animation.
What kinds of demos did Opus 5 produce?
The range spans games, explainers, and pure simulation toys:
A 3D art gallery let the user walk through a space, look at generated artwork, and trigger sound effects on hover, essentially building an interactive Pokédex-style cataloging experience. A Kerbal Space Program-style orbital mechanics tool let the user launch an object and watch it fall into a stable orbit around a planet, including named orbital transfer concepts like a Hohmann transfer. A cloth simulation responded to being pulled and torn in a way the tester described as visually convincing for fabric dynamics. A zoomable fractal generator let the user adjust bloom, dispersion, and depth in real time. A cellular automaton sim let the user drop “sand” and trigger effects like oil spills, steam, and fire inside a tube-shaped environment. A quadcopter flight simulator featured procedurally generated terrain, audio feedback from the drone, and functional landing controls. A pixel art editor replicated basic Microsoft Paint functionality, including layers, onion-skinning, and frame-by-frame animation for building simple movies. A population explainer graph visualized city population changes over 126 years using live, draggable data. A wrecking ball simulator let the user swing a ball into a brick and glass structure, with adjustable mass and a real-time destruction percentage.
None of this required iterative back-and-forth prompting in the demos shown. The tester’s framing was that these outputs, not long ago, would have been considered standalone games or applications in their own right.
Is Opus 5 worth using over other frontier models?
Based on the benchmarks referenced during testing, Opus 5 outperforms its predecessor Opus 4.8 broadly and edges out competing frontier models on several agentic and reasoning tasks, while costing meaningfully less per task.
Specific figures cited: Opus 5 scored 90.8% on BrowseComp (agentic search), roughly in line with GPT-5.6 and slightly ahead of a comparably-named rival model. On GDPval (knowledge work), it scored 1861, putting it above average human performance on the tasks measured. On ARC-AGI-3, a benchmark for novel problem solving, it scored 30.2%, compared to about 2% for both Opus 4.8 and GPT-5.6. On a new Automation Bench measuring business workflow generation, it scored 26% versus 17.4% for a rival model. On a “Humanity’s Last Exam” style benchmark, Opus 5’s lower-effort setting scored around 56% and its highest-effort setting scored close to 65%, at an average cost of about $3 per task.
Cost efficiency is where the gap seems widest. On agentic computer-use tasks, Opus 5 at a low reasoning effort scored about 60% for roughly $9 per task, and at high effort scored about 70% for around $22. A comparably capable rival model needed nearly $50 to hit similar scores to Opus 5’s mid-tier setting. Across most benchmark categories, Opus 5’s cost-per-task curve clustered tighter and lower than competitors, meaning less variance in both price and performance regardless of how much reasoning effort was dialed in.
It’s not a clean sweep. On agentic coding specifically, GPT-5.6 reportedly still scored higher than Opus 5, and legal-domain performance (11.7%) remains low across the board, reflecting how far frontier models still are from reliable specialized professional work.
What are the practical limits of these demos?
The showcased outputs are impressive as one-shot generations, but they lean toward visual and simulation-style tasks rather than production software with edge-case handling, security review, or long-term maintainability. A double pendulum demo, for instance, had a minor rendering glitch in one corner, the only visible error across the demos shown. Automation Bench scores around 26% indicate the model is still wrong roughly three out of four times on real business workflow automation tasks, even though that represents a sizable jump over the prior generation’s 17.4%.
The takeaway for builders is that Opus 5 is strong at generating self-contained, visually rich, interactive artifacts quickly and cheaply, but benchmark data suggests it should still be checked closely on tasks involving business logic, legal reasoning, or long-horizon agentic coding, where competing models currently hold an edge or where absolute performance remains modest.
Frequently Asked Questions
What is Claude Opus 5?
Claude Opus 5 is Anthropic’s frontier large language model, positioned as a significant step up from its predecessor Opus 4.8, with strong results on agentic coding, reasoning, and simulation-style creative coding tasks.
Can Opus 5 really build a 3D game from one prompt?
Demonstrated examples include a walkable 3D art gallery, an orbital launch simulator, and a drone flight sim, all generated from single prompts with minimal follow-up editing, according to hands-on testing.
How does Opus 5 compare in cost to other frontier models?
In testing, Opus 5 matched or beat a comparable rival model’s benchmark scores while costing substantially less per task, in some cases roughly a fifth of the price on agentic computer-use tasks.
Is Opus 5 better than GPT-5.6?
It depends on the task. Opus 5 scored higher on ARC-AGI-3 and Automation Bench, similar on BrowseComp, but GPT-5.6 reportedly still led on agentic coding benchmarks specifically.
What are Opus 5’s weak points?
Legal reasoning tasks scored low (11.7%), Automation Bench performance, while improved, still sits around 26% (meaning frequent errors), and agentic coding trails at least one competing model.