I have 200MAX subscription for Anthropic and a RTX 5070TI 16gb GPU. I only use Claude.AI web chat and I only use Opus4.7. I build my prompts on prem with Qwen3.8:latest and copy paste them in claude.ai. I never type anything into claude.ai I am merely a copy pasting monkey. If I need to adjust something I do it from the the on prem prompt. I download code bundles from Claude.AI and push them to my git repo I host. I instruct claude to not do any testing and only let me do testing on my machine so i am not wasting claude resources. The point is to minimize the claude agents from doing anything but specifically doing coding. I can do this for 12 hours a day and reach about 80% of my weekly usage. Once I get to 3 hours left on my weekly I start a fresh chat session with Fable5 and have it do blind code reviews on everything. I only do this because I want to max out my weekly allotment and fable will get me to 20% in 3 hours on a 200MAX sub. Sometimes Fable5 does something interesting but for the most part my on prem static and dynamic code analysis have kept everything buttoned up.