Claude Opus 4.8 vs GPT 5.5 Codex in Blender: who really builds better 3D scenes?

What happens when you give two cutting-edge AI models direct control over Blender and ask them to build full 3D scenes from scratch? In this showdown, Claude Opus 4.8 and GPT 5.5 (via Codex) go head-to-head on three challenges: a Pokéball, SpaceX’s Starship on the launch tower, and a fully detailed dragon. Along the way, you’ll also see how to connect Blender to AI agents using the MCP (Model Context Protocol) and how to bring AI into your video editing workflow.

What is Blender MCP and why it matters

Blender MCP is an MCP server that exposes Blender as a tool an AI agent can control. Instead of just generating code or giving you step-by-step instructions, the AI can actually manipulate objects, materials, lighting, and cameras inside Blender in real time.

With MCP, Blender becomes part of the AI’s toolset. The model can:

• Create and edit meshes
• Adjust materials, shaders, and lighting
• Set up cameras and animations
• Render stills and turntables

That’s what powers this entire comparison: both Claude Opus 4.8 and GPT 5.5 Codex are talking directly to Blender through MCP, not just telling a human what to click.

How to connect Blender MCP to your AI agent

Before you can run similar tests yourself, you need to connect Blender MCP to your AI of choice. The setup is straightforward once you know the steps.

Installing and enabling Blender MCP

First, get the MCP server running on your machine:

1. Open your terminal and run: brew install uv (this installs the UV tool used to run MCP servers).
2. Go to the Blender MCP GitHub repository and download the Blender add-on file.
3. Open Blender and go to Edit → Preferences → Add-ons.
4. Click Install from Disk, select the downloaded add-on file, and enable it.
5. In the 3D viewport, press N to open the side panel.
6. Find the Blender MCP section and click Start MCP Server.

Once the server is running, Blender is ready for your AI agent to connect over MCP.

Connecting Claude (desktop) to Blender MCP

If you’re using Claude’s desktop app with code tools enabled, you can hook it into Blender MCP like this:

1. Open Claude desktop settings.
2. Go to the Developer section and click Edit config.
3. In the claude_desktop_config.json file, add a new entry under mcpServers for Blender, with:
• command set to uvx
• args including blender-mcp (or the appropriate package name)

After you save the config and restart Claude, the Blender tools will load automatically, and Claude can start issuing Blender commands through MCP.

Connecting GPT 5.5 Codex to Blender MCP

Codex (the GPT 5.5-based app used here) has built-in support for MCP servers:

1. Open Codex settings.
2. Go to the MCP Servers section.
3. Click Add server and fill in the configuration to point to your Blender MCP server (matching the host, port, and command setup).
4. Click Save and reset the Codex app.

When you start a new session, Codex automatically detects the Blender MCP server and exposes Blender tools to the model.

Round 1: Pokéball – simple shape, subtle details

The first challenge is deceptively simple: build a Pokéball. It’s just a sphere, but the proportions, band placement, button, and materials all matter.

Claude Opus 4.8’s Pokéball

Claude Opus 4.8 starts by blocking out a sphere, splitting it into top and bottom shells, then adding the black band and center button. The full build takes about 7.5 minutes.

The result:

• Proportions are accurate and match the reference well.
• The button and chrome ring are clean and well-placed.
• The black band sits correctly around the midline.
• The red material is glossy and candy-like, slightly coral and more reflective than the flatter, brighter red of the reference.

Opus even self-critiques the color, noting the red is a bit desaturated and could be punchier. Overall, it delivers a polished, faithful Pokéball in a short time.

GPT 5.5 Codex’s Pokéball

Codex takes a more methodical, feature-rich approach and spends about 12.5 minutes on the build.

Its Pokéball includes:

• Separate beveled shells for added realism
• A recessed, satin-like black band
• Procedural scuff marks for subtle wear
• A full turntable animation and a display plinth
• A more muted, realistic red and a button with a soft cyan glow

Against the reference, Codex nails the proportions and the matte finish actually feels closer to a real toy. Where Opus goes glossy and punchy, Codex leans refined and restrained.

Verdict for Round 1: both are strong. Opus wins on speed and punchy visuals; Codex wins on realism and extra touches like animation and surface wear.

Round 2: SpaceX Starship – scale, structure, and smoke

The second challenge raises the difficulty significantly: SpaceX’s Starship on the launch tower, complete with booster, tower lattice, and venting plumes. This is a complex scene with lots of structural detail and atmospheric effects.

Claude Opus 4.8’s Starship attempt

Opus starts by blocking out the tower and stacking the booster and ship. However, this build runs into technical friction:

• Renders time out multiple times.
• The MCP connection drops and reconnects during the process.
• Lighting issues leave the final shot dark and underexposed.

Despite the problems, Opus manages to rough in the full tower and rocket. But because of errors before the final render resolves, the result is a murky, hard-to-read image. The total time balloons to around 26 minutes.

The underlying structure is there, but the presentation and reliability fall short.

GPT 5.5 Codex’s Starship

Codex approaches the same scene with a more stable, methodical workflow:

• It constructs the launch tower, booster, and ship with clear, readable proportions.
• It adds venting plumes and environmental details.
• It even includes text like “Gateway to Mars” on the pad.

The entire build completes in about 10 minutes, far faster than Opus’s troubled attempt. Side-by-side with the reference image, Codex’s version stands out:

• The silhouette of the rocket and tower is convincing.
• The venting and smoke help sell the scale and drama.
• The tower is simplified but still captures the essence of the structure.

Verdict for Round 2: Codex clearly wins. It’s faster, more stable, and produces a much closer match to the Starship reference.

Round 3: Dragon – the ultimate modeling test

The final challenge is the most demanding: a fully detailed dragon with wings, horns, scales, and glowing accents. This pushes both AI agents into more creative, sculptural territory.

Switching Claude to Ultra Code in the terminal

For the dragon, the setup switches from Claude desktop to Claude Code running in the terminal, using a newer Ultra Code backend. This requires a separate MCP configuration from the desktop app.

To simplify this, a “super command” script is used: you paste a single long command into your terminal, which handles the configuration and launches Claude connected to Blender MCP. Once it’s running, Ultra Code can drive Blender just like the desktop version, but with upgraded code capabilities.

Claude (Ultra Code) dragon

With Ultra Code powering Claude, the dragon build is ambitious and detailed:

• The body is sculpted with a strong, dynamic pose.
• The neck and tail are extruded and shaped for character.
• Large membrane wings are modeled with clear structure.
• Glowing blue horns and spine accents give a striking, icy look.
• Studio lighting and a turntable animation are set up to show the model from all angles.

Compared to the reference, Claude captures the overall vibe very well—the icy glow, the wing membranes, and the armored body all feel on point. The main difference is style: Claude’s dragon is stockier and more stylized than the sleeker original.

The trade-off is time: this build takes about 43 minutes, making it the slowest project of the entire showdown. But the result is easily one of the most impressive.

GPT 5.5 Codex dragon

Codex tackles the dragon much faster, finishing in around 12.5 minutes. It blocks out the body, long neck, tail, and a large bat-like wing.

However, the final result doesn’t land nearly as well:

• The overall form is rough and less coherent.
• The resemblance to the reference is weak—only vaguely similar if you squint.
• The design feels underdeveloped compared to what Codex achieved on the Pokéball and Starship.

Verdict for Round 3: Claude (with Ultra Code) wins decisively on quality and style, even though it’s much slower. Codex is quicker but delivers a dragon that feels more like a rough draft than a finished creature.

Overall results: speed vs quality vs reliability

Across the three rounds, a clear pattern emerges:

Pokéball
• Claude Opus 4.8: Fast (7.5 minutes), glossy, and polished.
• GPT 5.5 Codex: Slower (12.5 minutes), more realistic materials, extra details like scuffs and animation.

Starship
• Claude Opus 4.8: Struggles with timeouts and dark renders; structure is there but presentation suffers.
• GPT 5.5 Codex: Stable, fast (10 minutes), and visually convincing, with strong proportions and smoke effects.

Dragon
• Claude (Ultra Code): Very slow (43 minutes) but highly detailed, stylish, and close to the reference’s mood.
• GPT 5.5 Codex: Quick (12.5 minutes) but underwhelming result; rough and only loosely similar to the reference.

In short:

• Codex shines on structured, architectural scenes like Starship and adds nice touches to simpler objects like the Pokéball.
• Claude, especially with Ultra Code, excels at complex, organic, and highly stylized models like the dragon, even if it takes longer.

If you’re interested in how these models compare in other real-world tasks, it’s worth checking out broader benchmarks like GPT 5.5 and ChatGPT Images 2 vs Claude Opus: real tests, real results or a focused breakdown such as Claude Opus 4.8 review: powerful, honest, but only a small step up.

Bonus: giving your AI full video editing powers with Video Edit MCP

The 3D scenes are only half the story. The final output—cuts, transitions, overlays, and timing—comes from video editing. That’s where Video Edit MCP comes in: it’s a video editing platform designed specifically for AI agents, exposing editing tools over MCP just like Blender MCP does for 3D.

Setting up Video Edit MCP

To connect your AI agent to Video Edit MCP:

1. Go to videoeditmcp.com.
2. Copy the provided curl command from the site.
3. Paste it into your terminal and run it to install the app.
4. Once the Video Edit MCP app opens, select your email and click Connect AI agent.
5. Confirm the connection; the app will show that it’s ready.
6. When you open your AI agent, the app finalizes the connection, and the agent gains full video editing capabilities.

From there, your AI can cut clips, arrange timelines, add transitions, and render final videos—completing the pipeline from AI-generated 3D scenes in Blender to fully edited video content.

Which AI should you use for Blender work?

Both Claude Opus 4.8 (and Ultra Code) and GPT 5.5 Codex can control Blender effectively through MCP, but they each have different strengths:

• Use Codex if you want fast, reliable results on structured scenes, product-style renders, and setups where stability and speed matter.
• Use Claude (especially with Ultra Code) if you care more about creative, organic, or highly stylized models and are willing to trade time for detail.

The good news is that with MCP, you don’t have to choose just one forever—you can wire up both, experiment on your own projects, and see which model fits your workflow and style best.