Voyager vs Claude Code with Blender: a blind test on 8 briefs

Give two agents the same model, the same prompt and the same copy of Blender. Does the setup around the model still change what comes out? We ran eight Blender briefs through Voyager and through plain Claude Code, both on Claude Opus 5.5, to find out.

Both agents built every scene in Blender, the free 3D suite, by writing Python and rendering headless. Unlike in our Remotion test, both sides got exactly the same words: "Use Blender." is part of every brief. So whatever differs below comes from the agent setup, not the model or the prompt.

Each brief shows the prompt both agents received and two results, A and B: five videos and three still images. Which side is which is shuffled for every visitor. Pick the one that answers the brief better, and once you have picked all eight, reveal the answers to see how often you chose Voyager.

Take the blind quiz

What we found

Our reviewer judged the same eight pairs blind, and the split was clean. Voyager's result was preferred on all five briefs about real things: the mug, the logo sting, the bar chart, the earbuds and the reading nook. Plain Claude Code's was preferred on all three stylized briefs: the low-poly island, the isometric coffee shop and the glossy loop. That was one judge on one run per brief, so read it as a pattern to check, not a score.

The likeliest explanation is the skill. Every Voyager run loaded Voyager's Blender skill, a set of instructions for building, lighting, finishing and checking a Blender scene. Plain Claude Code had no Blender skill. That skill is written around real objects: materials, studio lighting, props that make a scene feel used, and a look at rendered frames before delivering. That shows in the details: the mug's speckled glaze and the coffee in the cup, the earbuds' mirror floor, the open book left on the nook's armchair. Its checking steps are written for real objects, though. On all three stylized briefs Voyager finished sooner than Claude Code, and on the coffee shop it delivered after a single preview while Claude Code kept fixing its scene. That is where Voyager's framing and colour choices slipped.

Time and cost went both ways. The mug shows the biggest gap: Voyager rendered it with Blender's fast Eevee engine and finished in 12 minutes, while Claude Code chose Cycles, Blender's path tracer, and took 48. Across all 48 runs (three per brief and side), Voyager averaged 7.8 minutes a run against 8.5, and cost more in model usage: $20.01 in total against $14.73.

What differs between the pairs

Both sides wrote their Blender scripts from scratch every time, and they often made the same technical choices: faceted geometry for the island, a true isometric camera for the coffee shop, clean loops on every video. Most of the differences are art direction. Voyager's results carry more set dressing and finishing detail; Claude Code's frame the subject larger and keep the scene simpler.

• Mug turntable. Voyager's mug has a speckled teal glaze over a bare clay foot, coffee inside and a two-step pedestal against a warm backdrop. Claude Code's has a single coral glaze, is empty and sits on a plain white pedestal against grey.
• Logo sting. The closest pair. Both start on an empty backdrop, extrude the Quillmere emblem, then bring in the wordmark and hold on the full logo. Voyager's backdrop is a cool light grey, and its colours come out paler than the logo file. Claude Code's is a warm cream with deeper shadows and keeps the logo's colours closer.
• Animated bar chart. Both read sales.csv and grow the bars one month at a time. Voyager's stands the bars on a dark, spotlit stage with a title, short labels such as $42.0K and no axis, and picks out June, the highest month, in orange. Claude Code's is a flat, light chart with a title and subtitle, a dollar axis, gridlines and full values such as $42,000.
• Earbuds hero shot. Both put a white case on a dark, reflective floor. Voyager's buds stand up in the case over a sharp mirror reflection, though that reflection shows the buds through the bottom of the case. Claude Code's case fills more of the frame, with the buds lying low in the tray and a soft reflection.
• Reading nook. Both rooms have the bookshelf, lamp, armchair and window under warm evening light. Voyager pulled back to show more of the room, with a green armchair, an open book left on its arm and a sky gradient in the window; both rooms put a mug on the side table. Claude Code framed closer, so its rust armchair is larger, with stronger sunlight falling across the rug.
• Low-poly island. Both islands are faceted, with palms and rocks, on a slow orbit. Claude Code's fills more of the frame and sits in a pale lagoon with haze. Voyager's is smaller, on a saturated blue sea, so much of the frame is open water.
• Isometric coffee shop. Both cut away the front walls of a small room. Both stock the counter with an espresso machine and a pastry case under a menu board. Voyager's is green and cream on a grass-edged base, with a cushioned bench, pendant lamps and a till. Claude Code's has a mint wall, two café tables beside a row of bar stools and a lit COFFEE sign that overlaps its menu board.
• Glossy loop. Claude Code's scatters many saturated shapes over a dark purple backdrop, some cut off at the frame edges. Voyager's builds around one large pink torus on a violet-to-coral gradient, so its main shape sits close to the colour of the background.

What to take from this if you use Claude Code with Blender

The model is the same on both sides, so most of what Voyager's skill adds is direction you can also write into a prompt yourself:

• Ask for the set dressing you want. In this test, plain Claude Code mostly built what the brief named and little more. If the mug should have coffee in it or the nook should look lived in, say so.
• Ask for framing in numbers. Voyager lost the island and the coffee shop partly on framing. "The subject fills about two thirds of the frame" is easy for an agent to check and hard to get by accident.
• Name the render engine and a time budget. The same mug took 12 minutes in Eevee and 48 in Cycles. For a stylized or preview render, "use Eevee" saves most of that.
• Give stylized work a contrast rule. A pink shape on a pink backdrop disappears. Ask for a backdrop clearly darker or lighter than the hero.
• Make it look before it delivers. Ask the agent to render three frames, describe what is wrong and fix it before the final render. All three briefs Voyager lost were ones where it spent less time than Claude Code.

Before the final render, render three frames (start, middle, end) and check them: the subject fills about two thirds of the frame, it stands out clearly from the background, and every element the brief asks for is visible. Fix anything that fails, then render.

To let Claude Code work inside a running Blender instead of from the command line, see Blender MCP: official vs community server and setup.

How the test was run

• The same model. Both sides ran Claude Opus 5.5 (claude-opus-5-5) at medium effort through Claude Code 2.1.301.
• What "Voyager" means here. Voyager's agent setup from Voyager build 2026.10.15+8: its system instructions, its skills, including its Blender skill, and the voyager command-line tool. It made its own choices instead of asking questions.
• What "plain Claude Code" means here. Claude Code with its default tools and nothing added: an empty home folder with no plugins, added skills, instruction files or connectors, and no voyager tool.
• The same prompt, word for word. Each brief is shown in the quiz exactly as both sides received it. Files a prompt names (the logo and the sales data) were copied into both project folders.
• Both used Blender. Every run on both sides built its scene in Blender 5.2.2 LTS with Python and rendered it headless.
• Same machine, one run at a time. Every run used one Apple M5 laptop with a 60-minute limit, and the side that went first alternated. Every run finished inside the limit and delivered a file.
• A fresh, empty project folder for every run.
• No paid media. Neither side could use image, video or 3D model generation services, so every frame was rendered on the laptop.
• Not cherry-picked. We ran each brief three times on each side, and the quiz shows run 1 of each, a rule fixed before the seven new briefs ran. The mug was our pilot and ran first, so its run 1 is the pilot pair.
• Shown at a common size. Videos are silent re-encodes at 720p; none had a sound track. Still images are scaled to 1280 pixels on the long side.

Frequently asked questions

Is Voyager better than Claude Code for 3D work in Blender? On these briefs it depended on the kind of brief: Voyager's results were preferred on the real-world subjects and Claude Code's on the stylized ones. A separate, larger internal test of ours (six harder Blender production tasks, repeated three times per setup, judged blind by two AI judges) also found no reliable overall difference. Take the quiz and see where you land.

Does Voyager use Claude? In this test, yes. Both sides ran the same Claude model through Claude Code. What Voyager added was its instructions, its skills and the voyager command-line tool, so the comparison holds the model fixed and varies the agent setup.

Were the two sides set up the same apart from Voyager? Almost. Voyager's side used Voyager's own bundled ffmpeg, while Claude Code used the copy installed on the laptop. Voyager also ran signed out and without the parts that only exist inside its app, such as its session tools, so nothing paid or networked was available to it.

Did anything change during the test? Two things. The mug pilot started with a 45-minute limit, raised to 60 while Claude Code was still rendering; it finished in 48 minutes, so under the original limit it would have failed, and the pilot did not alternate which side went first. And both bar chart runs were discarded and run again after the laptop went to sleep mid-run.

How do I use Claude Code with Blender myself? Claude Code can drive Blender from the command line with Python, as both sides did here; the tips above are a good start. To let it work inside a running Blender, see Blender MCP: official vs community server and setup.