GPT-6 Astra
medium effortScore8.8/10
AI Performance Tracker
A kappa takes a motorcycle for a spin. Codex and Pi redraw it every hour. Inspired by Simon Willison
GPT-6 Astra
medium effortScore8.4/10
GPT-5.6 Sol
high effortScore8.1/10
GPT-5.6 Sol
high effortScore7.3/10
GPT-6 Luna
medium effort- Vehicle mismatch
Score6.3/10
GPT-5.6 Sol
medium effort- Missing wheel
Score5/10
GPT-5.6 Sol
medium effortScore7.1/10
Loading older results…
About this AI drawing benchmark
One task, repeated across hours
Codex and Pi are scheduled hourly across 20 model, effort and agent configurations for each style. Both SVG and pixel-style artwork are generated as SVG from text. The timestamp on each card is the actual generation time; an hourly group may still be filling in.
The SVG prompt
Generate an SVG of a kappa riding a motorcycle through a desert, with spinning wheels.
How the drawing score works
Grok first describes the image, then assesses seven dimensions from 0 to 10. The weights below add up to 100%. We multiply each dimension score by its weight, add the results and round to one decimal. The final score is out of 10.
Example from the results above: GPT-6 Astra ·
| Dimension | Weight | Example / 10 |
|---|---|---|
| Character | 15% | 9 |
| Motorcycle | 15% | 9 |
| Desert setting | 10% | 9 |
| Riding pose and contact | 20% | 8 |
| Shape and construction | 15% | 9 |
| Composition and craft | 15% | 9 |
| Overall impression | 10% | 9 |
| Weighted result | 8.8/10 | |
A bicycle instead of a motorcycle is marked as a task mismatch. Its resemblance affects the relevant dimensions; it does not impose a separate score cap. Wheel motion is checked separately from drawing quality. Incomplete assessments stay unscored.
What to compare
Keep the model, effort, agent and style the same when comparing hours. A drawing score is evidence about this task. Look for repeated changes and check them against your own work before deciding when to use a model.