How Claude Opus 5.5 is quietly becoming a 3D and Blender powerhouse

Only a year ago, using an AI model to build complex scenes in Blender felt like a fragile experiment. Today, the latest Claude Opus 5.5 models are building castles, animating characters, simulating physics, and even creating playable mini-games—often with production-style structure and clean scenes.

Looking at the same tests run a year apart is one of the clearest ways to see how fast AI is advancing. The difference between earlier Claude Opus 4.x models and Opus 5.5 inside Blender is nothing short of wild.

From first castle attempts to procedural worlds

The journey started with a simple challenge: create a castle on a basic landscape in Blender, controlled entirely by an AI model. A year ago, using an early Claude 3 / Opus 4.x variant, the result was already impressive for its time—recognizable geometry, a usable scene, and a castle that at least looked the part.

Running the same test with Opus 5.5 today is a completely different story. The new model builds a far more detailed, coherent scene, with complex procedural structures and far better use of Blender’s tools. The castle isn’t just a rough blockout; it looks like something you could actually start lighting, texturing, and rendering for a real project.

This kind of test isn’t cheap yet: the castle scene took around 72 minutes of agent time and would cost roughly $50 in API usage if run purely through paid calls. But the quality leap in just one year is obvious.

The dragon and the pot of gold: a classic prompt, reinvented

One of the earliest public Blender–LLM experiments used a simple but fun prompt: create a scene with a dragon next to a pot of gold. Back then, the result was enough to go viral because it proved that an AI agent could drive Blender at all.

Feeding the same prompt to Claude Opus 5.5 today produces a scene that looks like it came from a completely different era. The dragon, the environment, and the composition are all built procedurally, with far more detail and structure. Rendered out, the scene looks polished and intentional rather than like a tech demo.

The gap between the two versions—separated by only a year—feels less like an incremental update and more like an entire generation jump in capability.

The donut test: has AI “solved” beginner 3D scenes?

If you’ve touched Blender, you know the famous donut tutorial. It’s become a benchmark for 3D beginners—and now, unintentionally, for AI models too.

Earlier tests with Opus 4.7 produced decent donuts: recognizable shapes, okay materials, a passable result. But they still looked like an early student project.

With Opus 5.5, the same simple prompt—“make me a donut in Blender”—now produces a scene that looks shockingly close to a human-made tutorial result. Not only does it create a realistic donut, it also sets up a cup and an extra chocolate donut, builds a clean Blender scene, and wires everything together procedurally. For many artists, this is now beyond what they could quickly build from scratch.

In a second round, the model was given a reference image: the classic Blender Guru donut. Opus 5.5 rebuilt a scene that is almost identical to the reference. At first glance, it’s easy to think the image was somehow reused, but it’s actually a full 3D scene constructed around the reference, with nearly 1:1 composition and lighting.

This two-round donut test took about 100 minutes and would cost roughly $21 in API calls. The key takeaway: for simple hero objects and beginner-level scenes, AI is already frighteningly good.

How the tools and ecosystem make this possible

These results aren’t just about smarter language models. They also depend on the ecosystem around them: plugins, APIs, and 3D asset generators that an AI agent can orchestrate.

One example is the Blender MCP (Model Context Protocol) server and plugins that let an AI model directly operate Blender—creating objects, editing materials, running simulations, and more. On top of that, tools like the Hixville Blender plugin and Hixville API give agents access to model generation services such as Tripo, Meshy, Hunyuan, and others for 3D assets, materials, images, and even videos.

The result is a kind of AI production line: the model plans the scene, calls external APIs for assets, imports and places them in Blender, and then refines the layout, materials, and lighting. Opus 5.5 excels at coordinating this entire workflow end to end.

Animating a keyboard: from “pretty good” to complex motion

Another test focused on animation rather than modeling. The challenge: animate the keys of an old keyboard in Blender. Earlier with Opus 4.7, the AI produced a solid result—keys moved, the animation made sense, and it was already impressive that a language model could script this at all.

Running the same idea with Opus 5.5 pushes things much further. The new animation is more complex, more coordinated, and feels like something a motion designer might actually build. The AI sets up intricate timing and interactions, all within Blender, without manual keyframing by a human.

This keyboard animation took around 111 minutes and would cost about $23 via API, but the complexity of the result shows how quickly AI is learning to handle non-trivial motion.

Assembling an angel: advanced VFX-style animation

A more challenging test involved an angel model split into multiple parts, with the goal of creating a tornado-like effect that assembles the pieces into a full figure. Even with Opus 4.8, this took several attempts to get working, and the final result—while cool—was relatively straightforward.

Opus 5.5 turns the same concept into something closer to a full VFX sequence. Instead of a simple assembly, the model builds a complex system of visual effects inside Blender: swirling elements, layered motion, and a more cinematic feel. It’s not just animating; it’s designing a whole effect system.

This angel sequence took about 84 minutes and would cost around $18 in API usage. For a single prompt-driven run, that’s a significant amount of animation and effects work automated by an AI agent.

Rebuilding an island from an image

One of the toughest tasks for AI in 3D is spatial reasoning: understanding a reference image and rebuilding a matching scene in 3D space. An earlier test tried exactly this with an island composed of multiple objects. The AI had access to APIs to generate models, then it was asked to place them to match the image.

The first attempt months ago was only partially successful. Objects were rotated incorrectly, placed in the wrong spots, and the scene didn’t really line up with the reference.

Running the same test with Opus 5.5 produces a much closer match. The island layout is nearly identical, with only small flaws like missing or slightly misplaced stones. The big structures and overall composition are right where they should be.

This run took about 62 minutes, generating 13 Tripo 3D models, 18 image edits, and one custom material. Estimated API cost: around $20. While it’s not perfect, it shows that AI is getting better at understanding spatial relationships and rebuilding complex scenes from 2D references.

Character animation: waving, walking, running, jumping

Another classic limitation for AI agents has been character animation, especially when working with rigged 3D characters. Earlier experiments produced stiff, awkward movement that didn’t feel usable without heavy cleanup.

Using a free rigged character in Blender, Opus 5.5 was asked to create four basic animations: a wave, a walk, a run, and a jump. The results are still not perfect—but they’re a big step forward:

• Wave: Smooth and believable, clearly using the control rig properly.

• Run: A bit rigid but functional, with a clear, consistent gait.

• Walk: Lacks some natural dynamics, but structurally correct.

These animations are good enough to use as a starting point, especially if you’re comfortable refining keyframes. For more complex creatures—wolves, scorpions, and similar—AI-driven workflows are already usable, as long as you review and polish step by step.

This animation set took about 116 minutes and would cost around $22 via API. It’s not “press a button, ship a movie” yet, but it’s clearly moving in that direction.

Physics and Rube Goldberg machines in Blender

Early AI–Blender physics tests were simple: drop some cubes, press play, and watch a basic simulation run. At the time, getting an AI to set up even that much inside Blender felt impressive.

With Opus 5.5, the bar is much higher. When asked to “impress” with a physics setup, the model proposed and built a full Rube Goldberg machine on a wooden table. The resulting scene is packed with interacting elements—ramps, dominoes, moving parts—that appear to run as one or several layered simulations inside Blender.

Whether it’s a single continuous simulation or multiple chained setups, the outcome is clear: Opus 5.5 can now design and implement complex physics-driven contraptions rather than just dropping objects and hitting play.

Building a playable football game in minutes

Perhaps the most surprising benchmark is game creation. A previous AI-assisted football simulator project took about 45 minutes to build and produced a basic but interesting prototype.

Opus 5.5 was asked to create something better—still within Blender and game logic constraints. The result is a fully playable football game with:

• Procedurally created players

• An intro sequence and basic presentation

• Referees and in-game events like penalties

• Working controls (even if a bit tricky for non-FIFA players)

The game runs without obvious bugs during extended play sessions and feels like a genuine mini-experience, not just a tech demo. It took only about 23 minutes for the AI to create something close to the earlier, much more manual project.

If you’re interested in how AI models compare on similar practical tasks, it’s worth looking at broader benchmarks like this hands-on comparison of DeepSeek V4, Opus 4.7, and GPT 5.5, or tests that pit Claude against GPT in real builds such as GPT 5.5 and ChatGPT Images 2 vs Claude Opus.

What this means for 3D artists, developers, and creators

Across all these tests—castles, donuts, angels, islands, physics, character animation, and games—the pattern is the same: the jump from earlier Opus 4.x models to Opus 5.5 feels like a full generation leap, not a minor update.

For 3D professionals, this is both exciting and unsettling. On one hand, AI can now handle a lot of the boring setup work: blocking out scenes, wiring rigs, placing assets, and building first-pass animations or simulations. On the other, it’s clear that the gap between “prototype” and “usable” is shrinking fast.

Right now, AI is best used as a powerful assistant: let it generate the base scene, then step in as an artist or developer to refine, direct, and polish. The more complex and creative the project, the more your taste and judgment still matter.

But if this is what one year of progress looks like, the next round of tests will likely be even more shocking. Today’s “insane” results may soon feel like the new baseline.