What's this about?
Mollick had early access to Gemini 3.8 Flash. His verdict: a very good Flash model, but not equivalent to a frontier model. He demonstrates with the same shader prompt he has been running for months — a drowned neo-gothic tower — noting that 3.8 Flash produced its version very quickly, and pointing to Fable 5.1's version of the same prompt for comparison.
Why does the same prompt across models across months work as a review?
Because it holds everything constant except the model. A shader is a good choice specifically: it is code that produces an image, so you can see the quality of the reasoning rather than reading it, and it requires holding a visual goal and a mathematical implementation in mind at once. Running it on every release for months produces a comparison that no benchmark table conveys — you can watch the capability move.
Why is it worth noticing?
- "Not equivalent to a frontier model" is a useful corrective to a launch claiming to approach frontier performance.
- "Very quickly" is a real part of the verdict — Flash models compete on latency, and that is a genuine axis.
- Willison's pelican and Mollick's shader are now both long-running personal benchmarks, and both are more informative than most published evaluations.