Creative Benchmarks: Testing What AI Models Can Build
LMRank has always been about practical model evaluation - not just benchmark numbers, but what models can actually do. Today we're launching a new category that pushes this further: Creative HTML Generation benchmarks.
What We're Testing
Instead of measuring reasoning, math, or coding in the abstract, these benchmarks ask models to build complete, self-contained projects from a single prompt:
- Web3 Landing Page - Generate a full marketing site for a fictional crypto project, with hero sections, feature cards, tokenomics, roadmap, animations, and responsive design - all in one HTML file.
- 3D Safari Simulation - Build an interactive 3D experience using Three.js, with animals, terrain, camera controls, raycasting, and a day/night cycle - again, all in a single file.
How Scoring Works
Each output is scored on four dimensions:
- Visual Quality (30%) - Design aesthetics, color palette, typography, layout polish
- Interactivity (25-30%) - Animations, hover effects, camera controls, responsive behavior
- Completeness (25%) - All required elements present, nothing broken, works standalone
- Creativity (15-20%) - Original design choices, interesting visual metaphors, wow factor
Why This Matters
Most benchmarks test a model's ability to reason about code. These tests measure a model's ability to ship product. A high score means the model can take a vague product brief and turn it into something real - exactly what developers and founders need.
We're starting with two test suites and will expand over time. Every model in our catalog that's available on OpenRouter can be benchmarked. Results appear directly on each model's detail page alongside their overall score.
See It In Action
DeepSeek V4 was the first model through the pipeline, scoring 8.2/10 on the Web3 landing page and 7.8/10 on the 3D safari. Check out the results →
Want to run benchmarks yourself? The admin dashboard at /admin/ lets you fire off tests against any model. More test suites coming soon.
See also: Best Coding Models