I work on aicost.tools, and “concise Ruby cost more tokens than Rust” matches what we see: the count follows the tokenizer, not how short the code looks, and even the model changes it — Anthropic’s new Haiku 5.5 counts the same text as about 30% more tokens than Haiku 4.5. I’d also love to know whether the 112 attempts include the repair turns, because that’s usually where the bill grows. If you want to check your own snippets, we made a free in-browser token counter that shows the count per model: AI Token Counter — Tokens & Cost Across 165 Models · AI//COST