A 90% tool-level token saving can shrink to under 1% overall

An analysis argues that token-saving claims for AI agent tools depend entirely on what denominator you use. Studying 140 archived Codex CLI runs, the author found that shell command outputs eligible for a technique called RTK made up only 0.1618% of all tokens used across those tasks. Even assuming a perfect 90% reduction applied to every eligible case, the resulting saving on the total task cost worked out to just 0.96%.

In other words, a tool's advertised 'saves 90% of tokens' figure is often calculated only within the narrow slice that tool touches, not across the full agent trajectory that includes prompts, conversation history, reasoning, retries, and later rounds. The full analysis and raw data are published on the author's blog.