> it is was a bit worse than gpt 3.5 (at least in the same prompts that we already had)

I'm willing to believe that Gemini isn't as good, but my impression was that you expect a new model to not perform as well on your existing prompts because the training set and training methodology is different. That's why one of the major risks of an AI business is vendor lock in, because you spend so much time optimizing a prompt for a specific model and you'll have to redo much of that work in order to switch vendors.

That you gave up so quickly when trialing a new model suggests the problem is even worse than I thought—you're locked in to OpenAI because every other model will always look worse to you, even if it would be better if you took the time to tune a new prompt.