I use claude and codex inside bubblewrap sandbox. I have $20 subscription for both. I do not use IDE integration. My task for agents are mostly generating simple code, generating tests, code review and debugging of hard to setup issues. I mostly use weaker models on lower effort (sonnet or opus plan for anthropic and luna or sol for openAI). It feels to me that higher effort causes problems with alignment and better models are not worth the cost except for the code review. I am rarely out of tokens so I do not plan to pay more.