Articraft · gpt-5.6-sol vs gpt-5.6-luna
Cost, retries, and geometry across four object categories · 2026-08-11
1 · Setup
Articraft
is an agentic system that writes Python code against a CAD-like SDK to build articulated 3D
objects from a text prompt, compiles and validates the result, and exports a posable USDZ.
It is the supported successor to the original
mattzh72/articraft
harness. Its default OpenAI model slug gpt-5.6 resolves internally to
gpt-5.6-sol, the flagship tier; gpt-5.6-luna is the cheapest tier
in the same family.
Articraft ships with no bundled dataset — each run writes only to its own local
runs/ folder, with no retrieval or reference corpus. The only existing corpus of
generated objects is a separate repository,
Articraft-10K
(CC BY 4.0): 10,788 records across 246 categories, entirely self-generated by running the
original harness many times — not sourced from an external 3D asset dataset.
Four prompts, one per category, were each run once on gpt-5.6-sol and once on
gpt-5.6-luna through an internal OpenAI-compatible gateway (default reasoning
effort). All 8 runs are shown below as live, interactive viewers — drag a joint slider to
move it through its articulated range.
2 · Cost and reliability
| Category | Model | Attempts | Cost | Sol / luna |
|---|---|---|---|---|
| Furniture · folding step stool | gpt-5.6-sol | 3 | $2.82 | 23.8x |
| gpt-5.6-luna | 6 | $0.12 | ||
| Tool · needle-nose pliers | gpt-5.6-sol | 2 | $2.70 | 39.6x |
| gpt-5.6-luna | 5 | $0.07 | ||
| Robotics · 2-joint arm + gripper | gpt-5.6-sol | 2 | $1.92 | 24.5x |
| gpt-5.6-luna | 2 | $0.08 | ||
| Electronics · laptop clamshell | gpt-5.6-sol | 2 | $2.83 | 18.3x |
| gpt-5.6-luna | 5 | $0.15 | ||
| Total (4 categories) | 24.5x avg |
Sol total $10.27 · luna total $0.42 · robotics is the only category where sol and luna took the same number of attempts.
3 · Live comparison
Each panel is a live three.js viewer (client-side USDZ parsing) — drag a joint slider on the right of each panel.
4 · Image → articulated asset
The same pipeline also reconstructs from a reference photo: the agent views the image with its view_image tool, infers the object's shape and how it should articulate, then writes the CAD code. All four below succeeded end-to-end on gpt-5.6-sol. Left: input photo · right: live viewer — drag a joint slider to pose it.
5 · Takeaways
- Both tiers reached a passing compile on every one of the 4 prompts — luna is not a reliability compromise here, just a slower converger.
- Robotics was the closest race: identical attempt counts, same viable joint ranges, 24x cost gap.
- Furniture and electronics show the most visible detail gap between tiers (materials, part count) when opened in the viewers above.
- For a first pass or high-volume generation,
gpt-5.6-lunais the better default; reservegpt-5.6-solfor prompts that keep failing on luna. - The pipeline is not text-only: from a single reference photo (section 4), the agent reconstructs both the geometry and a plausible joint structure — e.g. it recovered the anglepoise lamp's coil springs and three revolute joints, and picked the right chair out of a cluttered store shelf. It writes CAD code from the image, not a neural mesh.