I really really really want Mistral to be my daily driver for (vibe) coding but it sucks. The difference is very clear when comparing to gpt-codex. night and day difference. I wanted to transform an app in one index.html into a modern vite/react app. Mistral vibe with devstral 2 struggled for over an hour, to just init the app. gpt-codex-5.2 oneshotted it in less than 10 minutes.
I tried to have the AI models calculate the amount of calories in a meal based on the ingredients and recipe. Mistral had wrong numbers for the calories in the ingredients, so the whole calculation was off. Gemini gave answers that were perfectly usable. A while ago I asked all AI models I might use the same question. The chat apps were Mistral, Gemini, Claude and ChatGPT. Mistral gave obviously incorrect answers, and failed to disambiguate a word that should've been obvious in context. The others were mostly correct, good enough to be usable.
Mistral did give answers that read the best, but that is because I gave Mistral and only Mistral a system prompt to be concise. When other models were given the same system prompt, their answers became nicer to read. I think there's a pretty obvious reason why Mistral sucks compared to the other models. It doesn't do any reasoning by default. If your agentic model doesn't do any reasoning, it can't do long tasks by itself. It can't handle tasks that have a few steps.