Developer Finds 97.6% of AI Coding Tokens Wasted on Logs and Test Output

A developer tracked 4,788 commands across AI coding tools Claude Code, Cursor, and Aider over seven days, logging 355 million total tokens consumed. Analysis revealed that 97.6% of those tokens — roughly 347 million — were spent on unnecessary output such as test results, build logs, package manager notices, and git status details. Test output alone accounted for 343 million wasted tokens, since tools like pytest generate thousands of tokens per run while only a handful of lines are relevant to the AI. At Claude Sonnet's pricing of $3 per million input tokens, the wasted tokens translated to over $1,000 in a single week for one developer. The author concludes that filtering terminal output before passing it to AI tools — while preserving stack traces in full — could cut costs by up to 40 times.