The quota question is the one that actually changed my behavior — I used to burn a weekly Claude window on trivial edits without noticing. The CP framing (score divided by cost per task) is the right metric too; most "which model is best" debates ignore the denominator entirely. My own addition to this stack: I keep a separate rotation of free LLM access points for non-coding work (freellm.net is the directory I check first), so the paid coding quotas stay reserved for actual coding. One report for five providers plus a free-tier rotation covers almost everything.