I am impressed by how fast Gemini 3.7 flash is, and it performs well enough.
I am a believer of harness before model. I use 9router to route Claude Desktop to use Gemini API because of the recent Deepseek's price hike. I know damn well how bad 3.5 flash and 3.6 flash were, but man 3.7 flash is actually good and it can understand my code-base immediately and was able to make a change gracefully and follow my instruction, rarely hallucinating and uses all the tools efficiently.
And it is fast. Really fucking fast. Unbelieveably fast. Most of the time complex tasks that would require v4 flash to complete in 10 minutes, 3.7 flash can finish them in 1-3 minute. And the I cannot see the output difference between the two models.