Model comparison

Evaluation

27 outputs from 27 models for the same zero-shot prompt – each one the model's first answer, published unedited.

• 4.4/8average score
• 20/27working outputs
• 86/100avg. technical health
• 56/100avg. UX

Broken (do not run): 7, 7 of them with a copy fixed by the same model. The average checklist score is 4.4/8 (a broken output scores 0), 6/8 among working ones. Best: Claude Opus 5.5, Claude Haiku 5.5, Claude Sonnet 5.5 and 3 more – 8/8. Weakest working: Gemini 3.1 Flash-Lite – 0.5/8.

How we evaluate – methodology →

How the 8 requirements were met

From 20 working outputs: ✓ met, ½ partly, ✗ missing. Rate = (✓ + ½ × 0.5) / number of outputs.

Strengths

The requirements the models met best – quotes from the reviewer's verdicts:

Single file, no external assets, full window, resize & high-DPI, runs without errors · 80 % met

• Single file, no assets, resize and high-DPI handled, renders fine. – Claude Opus 5.5
• Single self-contained file, resize and high-DPI handled. – Claude Haiku 5.5

3D scene · 78 % met

• Sun, eight planets, Moon, asteroid belt and stars all present. – Claude Opus 5.5
• Sun, eight planets, Moon, asteroid belt and starfield present. – Claude Haiku 5.5

Typical mistakes

The requirements the models most often missed or met only partly:

Camera · 68 % met

• Camera is coded, but the frozen page does not respond. – Mistral Medium 3.5
• Camera cannot be controlled. – Gemini 3.1 Flash-Lite

Orbit & label toggles, readable vs true scale switch · 73 % met

• Labels missing, orbits invisible, scale switch barely works – Mistral Medium 3.5
• Toggles do nothing; no scale switch. – Gemini 3.1 Flash-Lite

Click / list · 73 % met

• Selection and fly-to are coded but do not work on the frozen page. – Mistral Medium 3.5
• No body selection or info panel. – Gemini 3.1 Flash-Lite

Agreement of two independent reviewers

The checklist of 28 outputs was filled in independently by two AI reviewers, Claude Opus 5.5 and Codex (GPT-6 Astra, medium). They agreed on 186 of 224 items (83 %). Average score by Claude 4.8/8, by Codex 4.8/8. Codex gave a lower score for 12, a higher one for 8 and the same for 8 outputs. They disagreed most often on: Orbit & label toggles, readable vs true scale switch (10×), Click / list (7×), Camera (5×).

Where the reviewers disagree, the stricter (lower) verdict is proposed; the published score is approved by a human.

Objective tests

Measured by a tool, not by a person or an AI – so they are shown even before the scores are approved.

• Tech: average 86/100 from 27 measurements; highest GPT-6 Astra (100), lowest Mistral Small 4 (64). Measures Lighthouse, smoothness, errors during automated use, the mobile layout and whether the controls work.
• UX / UI: average 56/100 from 27 measurements; highest Grok 4.6 (94), lowest Gemini 3.1 Flash-Lite (0). Measures control size, layout at 3 widths, keyboard, feedback and readability.