It depends on the use-case. yes, 90% of cost is cache in agentic coding scenarios (actually 95% in my experience). But not when the model reasons for 200k+ tokens before answering a complex problem.