I gave Claude Opus 5.5 the exact same 3D conference prompt at Low, Medium, High, Extra, Max, and Ultracode effort.
I compare every output, along with the runtime, estimated API-equivalent cost, token use, verification checks, and questions asked. The most expensive run was not the best, and I break down which effort level I'd actually use for a complex creative build like this.