Start from your task, not a leaderboard. Collect 20 to 50 real examples from your work, run them through two or three models, and score the outputs yourself.
Often a smaller, cheaper model handles routine jobs like summaries, tagging and first drafts. Save the strongest model for hard reasoning or high-stakes output. Open models make sense when data must stay inside your environment, when you have high volume and want predictable costs, or when you need to customise. Count the hosting and maintenance effort too. Closed models are easier to start with and usually need no setup. Prices and rankings change often, so recheck current pricing pages and rerun your test set every few months.