Recently I upgraded to the $200 dollar plan and it wasn't a busy month for me so if I had to kill my usage limit I just set max on everything. I've been really curious about analyzing how to pick efforts myself instead of delegating the work out to an orchestrator, mostly just because I'm curious and wanted to see if I really need the $200 plan. Or maybe it'd be more beneficial for me to do lower effort fan outs. My investigation shows really statistics on token usage between xtra and max and other effort levels as well. It also tests to see when the higher effort was worth it. I call this a clean finish which is my stand-in for “it worked the first time.” It counts a prompt as clean when no shell command had to be retried, you didn’t interrupt, and your next message didn’t read like a correction. It doesn’t check whether the code was actually right. On the same kind of work, max generated 1.6x to 3.2x the tokens of xhigh. Max never finished cleanly more often than xhigh on any kind of work. Larger changes: 73% clean for max (75 prompts) vs 85% for xhigh (68 prompts). Small edits: 87% vs 97%. On small edits, high finished cleanly 94% of the time against xhigh’s 97%, for about half the tokens.

In the future I will be testing performance of fan outs at different effort levels. Anyways you can read more details on my blog https://danielyakobian.com/blog/claude-code-max-effort.html