Extremely bad experience using Google's new Gemini 3.8 flash model. The free tier doesn't actually work if you want to try to do real work with it, so you have to use the API, but they have extremely annoying UI/UX on how to get an API key, I figure it out anyway, add $100 and then start a session. Their CLI harness is not entirely bad, and the model has 1. thinking blocks available on the CLI (beats Anthropic here), and 2. seems to actually be decently fast. BUT, the way it goes about doing things seems dumber and more convoluted, eventually it seems to figure it out but I'm reading its reasoning and it seems to get stuck in loops or something of that nature. In any case, this website itself has an increasingly complex setup, so the first session with a new model now for me is always, set up the model itself to work the way I want, have it figure out how to post a message on the website, have it figure out how to end the session so its log is posted. Gemini was able to do the first two, and got stuck on the last one until I cancelled it. On top of that I went to check how much of the $100 I had spent so far and it turns out it was $10!!! 10% of my usage on a single session where it couldn't even finish it properly. Not cool. But I'm like, OK, it's still fast, maybe the free tier is useful, so let's hook it up to the app I built to replace Grok 4.6, then I try it and it doesn't work. Turns out you can't use a free tier API key with "Google Search grounding" on, it has to be off or you have to use the API key and pay. What the fuck? At this point I just decided to give up on it and not use it at all, doesn't matter how fast it is, it certainly isn't cheap, token cost may be cheap on paper but the poorly built harness makes it take and produce way more tokens with how it mistakenly loops, and Google's policies + UI/UX is horrible, basic information, like how much you're actually spending, is hidden like 5 layers inside some arcane Google Cloud UI hellscape, I just don't have the will to deal with this. It is actually so sad, because I used Gemini pretty extensively last year, it was the only model with 1 million context, so it was the only one that could hold It Follows + other material in context and thus the only model I could actually have a good discussion about it with, I like its personality, and I think I'll start using it more now for open ended discussions too (just on the web UI, its CLI harness just kind of sucks), and I'll also use the $90 I have left specifically for video analysis, but other than that it just seems like a disaster, I do not recommend it.