I think frontier models are bait now. I prefer a fast, less thoughtful model that adhere's to instructions. The code these latest models produce is often hot garbage and you still need to micro-manage. Fast with small chunks is better.
I've been messing with DS4 flash. It's very fast and has 1m context. Downside is it's probably not great for private codebases (at least not when hosted by DeepSeek) and it's also not great at sticking to instructions if the task is too large / long though. I expect it will improve over time with better tooling and cache usage.
I basically have it load up a bunch of relevant context and give it small chunks of work in the same session over time (not like a fire and forget subagent). It's working fairly well. Bonus is I still feel like I'm part of the process instead of watching youtube videos while Opus / GPT vibe code a bunch of slop.
