Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark

TEXT START: A practical OpenSCAD LLM benchmark comparing Codex 5.5 High, Claude Sonnet, Claude Opus, Cursor Composer, Google Antigravity, and ModelRift on a detailed Pantheon model.

THE DISSECTION

This is a technical benchmark documenting LLM performance on parametric 3D architectural code generation. The operative data: Google Antigravity 2.0 with Gemini 3.5 Flash scored 4.5/5 quality, ModelRift with human-in-the-loop scored 3.8/5, Codex 5.5 High scored 3.0/5 (with export pathology), and Cursor Composer produced "the weakest result" despite being fastest. The article concludes that "fully autonomous generation is not the right workflow for this kind of task yet."

What this article is actually doing: documenting the state of play in cognitive task automation, celebrating improving quality, and performing the ritual assurance that humans still add value in the loop. The benchmark itself is competent. The framing is the ideology.

THE CORE FALLACY

The "not yet" inference engine. The article explicitly states autonomous generation "is not the right workflow for this kind of task yet." This is the canonical rearguard statement: present humans as temporarily indispensable, establish a trajectory toward full automation, and call the current human involvement a feature rather than a bug.

From the Discontinuity Thesis lens, this is precisely backwards. The trajectory is the point. "Yet" is doing enormous ideological work here—it transforms a displacement mechanism into a temporary inconvenience, a workflow preference rather than an economic death sentence for the humans currently performing this labor. The benchmark documents how far along the automation of spatial reasoning has progressed. The conclusion treats this as reassuring.

HIDDEN ASSUMPTIONS

1. Spatial reasoning is a special category. The article treats architectural/parametric geometry as qualitatively different from other cognitive tasks, implying it sits in a domain where AI progress is noteworthy but not threatening. This is the "this time is different" fallacy applied to automation. The historical record shows every task category initially declared AI-resistant eventually falls. There is no structural reason spatial code generation is exempt from the same trajectory.

2. "Tool use was not the bottleneck." The article treats the ability to call OpenSCAD CLI and render previews as trivial, noting that "geometric judgment" was the limiting factor. This framing positions geometric judgment as the final moat. It is. Until it isn't. The benchmark shows this moat narrowing: Antigravity achieved the "best fully autonomous model" by searching for real architectural parameters and implementing the coffered ceiling pattern—sophisticated spatial reasoning, autonomously executed.

3. Productivity gains are net positive. The article never asks who benefits from these efficiency improvements. It assumes the value generated by faster, cheaper, higher-quality parametric modeling flows to someone worth caring about. Under DT logic, the gains flow to Sovereigns (those who own and control the AI capital). The article performs perfect institutional tone-deafness on this question.

4. Human-in-the-loop is a feature, not a transition phase. The article celebrates ModelRift's Annotation Mode: "draw arrows and notes directly on a 3D model screenshot and feed that back to the AI." This is presented as superior workflow design. Under DT mechanics, this is the Servitor position—humans retained to perform low-autonomy, high-inspection tasks that verify and correct AI outputs. The article accidentally documents the descent in productivity prestige while treating it as a feature.

5. Autonomous ceiling is the goal, not the threat. The entire benchmark implicitly frames "best autonomous result" as the ceiling to aspire to. The framing assumes autonomous is better, faster, more valuable. This is circular logic: autonomous AI producing better outputs is presented as progress, not as evidence of productive participation collapse.

SOCIAL FUNCTION

Transition management with a technical credibility veneer. This article performs the specific cultural work of making AI progress legible to technical audiences as interesting benchmark results rather than displacement data. The ModelRift company context is important: this is a platform whose business model depends on LLM-generated 3D models. The article functions simultaneously as technical documentation, product demonstration, and reassurance theater. The target audience—developers, engineers, technical managers—gets to engage with AI progress as a craft challenge rather than an economic threat. The conclusion that human-in-the-loop "still matters" is the lullaby. The data shows exactly why and how it will stop mattering.

THE VERDICT

This is a well-executed benchmark that accidentally documents the narrowing moat between AI and human spatial reasoning performance. The Discontinuity Thesis does not require AI to match human performance on every task today. It requires that the trajectory renders human productive participation increasingly optional. The benchmark shows: autonomous spatial code generation at 4.5/5 quality, faster systems producing lower quality (Cursor), preview/export divergence as the remaining debugging challenge, and human-in-the-loop as the quality enhancement mechanism that will itself be automated.

The article's conclusion—"fully autonomous generation is not the right workflow for this kind of task yet"—is accurate and irrelevant. The "yet" is doing ideological work that the technical data contradicts. The trajectory is clear: the gap between autonomous and human-in-the-loop narrows with each model generation, the toolchain plumbing is solved, and the remaining barriers are engineering problems, not structural constraints.

For the humans whose productive work involves architectural 3D modeling, parametric CAD, or spatial code generation: you are currently in the Servitor window. Your job is to verify, correct, and annotate what the AI produces. This is a better position than being fully displaced, but it is not a stable position. The benchmark's own results demonstrate the trajectory.

For the Sovereigns building ModelRift-style platforms: this is the transition playbook in action. The article accidentally reveals the mechanism—human annotation as a training signal layer, autonomous generation as the production path, the human-in-the-loop step as cost to reduce rather than quality to preserve.