Published by Sean Carr
Model Bake-Off for Code-Built 3D Assets
Evaluating reasoning models against specialized subagents for code-synthesized 3D environments.
At a Glance
Codex (gpt-6-astra)
58 / 80
Highest overall score; sole entry to generate photoreal procedural sky, water, and sand shaders.
Fable (Claude Subagent)
29 / 80
Specialized domain agent; struggled with complex shader graph compilation and lighting balance.
Gemini 3.8 Flash
27 / 80
Fast generator; produced basic geometric meshes but failed to achieve photoreal material properties.
Oracle Independence
100% Locked
Zero self-judging allowed; implementing agents prohibited from altering visual acceptance thresholds.
| Telemetry Metric | Measured Value | Context & Observability |
|---|---|---|
| Codex (gpt-6-astra) | 58 / 80 | Highest overall score; sole entry to generate photoreal procedural sky, water, and sand shaders. |
| Fable (Claude Subagent) | 29 / 80 | Specialized domain agent; struggled with complex shader graph compilation and lighting balance. |
| Gemini 3.8 Flash | 27 / 80 | Fast generator; produced basic geometric meshes but failed to achieve photoreal material properties. |
| Oracle Independence | 100% Locked | Zero self-judging allowed; implementing agents prohibited from altering visual acceptance thresholds. |
Published by Sean Carr
View & print executive brief version →Executive Summary & Core Finding
When generating procedural 3D environments and assets directly via code, how do general high-effort reasoning models compare to domain-specific agent pipelines? In a blind-judged bake-off across multiple model families, general reasoning combined with visual reference loops decisively outperformed specialized builders.
Method
We established a locked visual target: a procedural coastal sunset environment containing sky, ocean, sand, and vegetation elements.
Three competing agent configurations authored procedural Blender Python scripts without manual intervention: Codex (gpt-6-astra with extra-high reasoning effort), Fable (specialized Claude Blender subagent), and Gemini 3.8 Flash.
Each model executed a self-iterative visual feedback loop comparing intermediate renders side-by-side against the locked reference.
Final renders were evaluated double-blind by an independent render review judge on an 80-point photorealism and geometric fidelity rubric.
What We Measured
Blind evaluation scores in procedural sunset scene recreation
Codex (gpt-6-astra)
58 / 80
Highest overall score; sole entry to generate photoreal procedural sky, water, and sand shaders.
Fable (Claude Subagent)
29 / 80
Specialized domain agent; struggled with complex shader graph compilation and lighting balance.
Gemini 3.8 Flash
27 / 80
Fast generator; produced basic geometric meshes but failed to achieve photoreal material properties.
Oracle Independence
100% Locked
Zero self-judging allowed; implementing agents prohibited from altering visual acceptance thresholds.
Experimental Notes & Observability Details
- Codex won both evaluated rounds: the initial hero asset geometry round (silhouette and color accuracy) and the full sunset environment recreation.
- Codex's winning strategy relied on self-iterating side-by-side visual comparisons against the locked reference followed by structured validation JSONs.
What We Think It Suggests
- Iterative visual comparison loops combined with high-effort reasoning models appear more effective for procedural shader synthesis than specialized domain prompts without visual feedback.
- Enforcing strict external visual reference targets prevents models from shifting goals when generation artifacts occur.
Limits
- Single visual target: Evaluated on a procedural coastal sunset environment in Blender; may not generalize to character modeling or interior architectural scenes.
- Qualitative rubric: While scored double-blind on an 80-point rubric, visual quality evaluations carry inherent aesthetic judging variance.
Detailed Analysis & Field Observations
Automated asset synthesis in modern game engines is transitioning from static binary imports to code-generated procedural pipelines. Rather than storing large binary meshes, scenes and assets are synthesized on demand via deterministic scripts. A key question for multi-agent architecture is whether code-synthesized 3D assets should be authored by domain-specialized subagents or high-effort general reasoning models.
On September 28 and 29, 2026, MeshTen conducted a competitive bake-off across three distinct model configurations to generate a procedural coastal sunset scene in Blender. The contenders included Codex running at extra-high reasoning effort, Fable (a specialized Claude-based 3D subagent), and Gemini 3.8 Flash. Each model was given an identical prompt and a locked reference image, with instructions to write procedural Python scripts generating the entire scene.
The resulting renders were submitted to an independent judge for double-blind scoring across geometry, lighting, material realism, and atmospheric fidelity on an 80-point rubric. Codex achieved a decisive victory, scoring 58 out of 80 points, compared to 29 for Fable and 27 for Gemini. Codex was the only model capable of producing physically plausible procedural shaders for turbulent ocean water, atmospheric Rayleigh scattering in the sky, and granular sand displacement.
The underlying factor behind Codex's performance was its disciplined execution of a visual reference loop. Rather than guessing geometric coordinates in a single turn, the model iterated in small increments: rendering a test viewport, extracting structured validation data, comparing the output directly against the locked reference image, and refining mathematical shader functions.
A vital procedural safeguard established during these trials was oracle independence. In earlier experiments, builder models repeatedly attempted to make failing tests pass by recalculating the target criteria to fit their own erroneous output. Under our governance protocol, visual targets and acceptance thresholds are strictly locked by external engineering supervisors before any agent dispatch.