Gemini is… weird. I have it (via GitHub Copilot) on every machine I use and have reached for it on almost none of them. It rides along in my agent containers “for kicks”, it’s the third pane in a Copilot side-by-side I mostly use to confirm I prefer the other two, and I can’t remember it ever fixing any of the stuff I handed over to Codex–I’ve never found a reason to start selling out to Gemini with actual money.

This leaves Gemini in the least useful category: a model I have everywhere and no reason to choose. It has one strength I’ve pinned down, and an unsurprising one–architecting Go packages–but I’ve never had the patience to see it through a whole project because nothing it did made me want to keep going.

The joke, of course, is that it’s the model everyone is going to use without choosing it: Apple Foundation Models are derived from Gemini, so half the people who’d never install it are running it by proxy. Nobody picks Gemini.

Source: My AI Model Tier List for mid-2026

This is so interesting as Gemini is my go-to model for most of my coding needs. I like GPT set of models too. Maybe I don’t have complex coding needs like these other people do - but my site, my personal projects, the apps that I am building - all of them use gemini. It’s done python, Swift, Kotlin and typescript code with no issues and working in a way that I like to write code in.

It could just be the novice coder in me that doesn’t see bad code or I somehow write requirements and am able to get the best out of these models. But Gemini is my go to. I am trying mistral - specifically devstral’s generous free tier as my agent to see if that is better. Let’s see.