Turning thinking off cut my API bill 24% and the time to summarize 20 podcast excerpts by 42%. Artificial Analysis puts Haiku 5.5 output at ~162k tokens per task, roughly 3x GPT-6 Luna. My run showed the same thing. Across 20 excerpts, Haiku charged 22,684 output tokens vs Luna's 10,581 (2.1x), mostly hidden thinking. Two other settings set the bill. A 1,000-token output limit left 17 of 20 answers cut off. You're still billed for what's cut off. Once a single prompt passes 100K tokens,…