The average person’s “model comparisons” today are really “prompt comparisons.” API behavior ≠ product behavior. Gemini 3 in Antigravity, Cursor, Gemini App isn’t the same as what you’d get via the API. Same with o3, Sonnet, Opus, etc. I always spend a little time using the model in the playgrounds to get a sense of their capabilities, and then inside of products to see the opinions of the product teams wrapping it. The wrapper matters more than people think. The model is still raw clay; product decides the form.