Opus 4.7 Made Me Take Token Waste Management Seriously

Anthropic shipped Claude Opus 4.7 on April 16, 2026. Same per-token price as 4.6. New tokenizer. The official docs say it quietly: "This new tokenizer may use up to 35% more tokens for the same fixed text."

That line is what finally made me stop hand-waving about AI costs and actually audit where my tokens were going. I run Claude agents all day across my stack, and a silent 35% hike on the exact same prompts meant the ceiling for "good enough" just got a lot lower.

Here's what's in the full post:

• The two ways tokens leak — I separate waste (turns that produced nothing useful) from inefficient usage (turns that worked but cost way more than they should have). Most guides collapse these into one bucket. They shouldn't.

• How I measured it across 133,087 turns — I built a token waste sorter over 9,667 sessions and ran model-vs-model comparisons for the judgment calls. I explain the methodology, what I kept, what I threw out, and which clusters dominated the bleeding.

• The top waste cluster was not what I expected — It wasn't hallucinations, bad reasoning, or runaway loops. It was infrastructure. Browser/Playwright failures spread thin across hundreds of cheap sessions outweighed every "smart" optimization I'd been chasing.

• Three fixes that cost nothing — Shrinking CLAUDE.md, setting tight max_tokens, and auditing WebFetch failures. I share the before/after for each and which one moved the needle hardest.

The part that surprised me most is in the full post: the meta lesson about where budget actually goes when you measure instead of guess.

Full post: https://thoughts.jock.pl/p/token-waste-management-opus-47-2026

Free weekly: https://thoughts.jock.pl