As others have noted, Gemini Flash 002 is much worse than 001.
I created an app for the Gemini Competition (recime.ai), and noticed that it recently started throwing errors.
After digging into it, I discovered Google had changed the default “gemini-1.5-flash” model from the “001” version to the “002” version.
Previously with Flash 001, the responses returned from the model were almost always perfectly formatted and structured, which made for easier parsing.
Now, the responses returned from this newer 002 version are almost always incorrect.
Furthermore, the newly incorrectly returned structure / format was never returned by the 001 model, meaning any previously painstakingly implemented edge case error handling did not work with this new 002 model.
Different types of prompts with 002 do not seem to make it behave like we want it to (whereas 001 was more willing to listen to your prompts).
Sadly, it looks like the smarter Flash 001 model will be discontinued in May 2025.
Did Google do a cheeky and swap in a cheaper model after the Google Gemini API Competition hype died down?