Does Gemini 2.5 Flash use different models for reasoning vs. non-reasoning outputs?

I noticed that the cost of reasoning outputs is significantly higher—around 6 times more—compared to non-reasoning outputs (3.5 vs 0.6).

If both use the same underlying model, wouldn’t using <thinking> tags with reasoning turned off produce similar results at a lower cost?

Has anyone tested or analyzed how Gemini 2.5 Flash behaves in this kind of setup?