I've only tried Claude Opus but it does a pretty good job interpreting the plan. It's also really powerful being able to bring in telemetry context and code to help make a more balanced tradeoff with usage metrics (much easier to callout anomolies, edge cases, and non performance sensitive contexts)
That said fixing the queries which usually originate from Django ORM is a bit hit or miss. Been multiple times the LLM wants to rewrite to something similar but not identical or wants to make large, structural changes instead of focusing on low hanging surgical improvements
Once it was even able to pull PG mailing lists to figure out I was hitting a planner predicate comparison limitation where switching from "bool = false" to "not bool" or maybe the other way around led to a massive 100x+ improvement (the planner was skipping using an index because the predicate in the query didn't "match" the predicate in the partial index despite the conditions being logically equivalent)
Would have taken days or weeks for me to figure that out on my own