I've been using Opus and Gemini. Opus used for the final design/architecture reviews and some of the more complex RCA fixes. Gemini Argon for everything else including code gen. I found this was the best way to cost optimize token usage vs performance.

I had tried open models and a few others including Astra/Sol. But for my usage the best option was Opus + Gemini.

Gemini goes a long way with low token usage but delivers very good performance. My primary use case right now is building an application that integrates generative CAD + CFD + DFM optimization that's specifically aimed for UAV design (airframe, internal modules/bays, etc).

Edit:

I also wanted to add a bit about my worflow/processes because I think that really matters in which models work well for you. I do a lot of the architecture and design work up front. I have detailed FDDs, TDDs, RCA docs, implementation plans, etc which then get fed to Gemin to do the code changes. I basically treat Gemini like a junior dev. I think if you're trying to do one shot prompt applications, then another model (ie. Fable) will likely be a lot better.