Your tenth message costs far more than your first. Grab the guide + the Token Saver skill: clean tasks, evidence selection, settled steps as code, accepted answers reused.

Source: Nate’s Newsletter Date: 2026-07-29 URL: https://natesnewsletter.substack.com/p/reduce-ai-token-usage

Summary

Same underlying piece as Nate’s Newsletter’s “Token Saver” post: the author measured that 95.73% of their AI input tokens (out of 3.77 billion) were reused context rather than new material, meaning a session’s tenth message can carry far more token weight than its first even when what’s typed is shorter. The accompanying guide and “Token Saver” skill frame the fix around four habits — clean tasks, deliberate evidence selection, treating settled steps as code rather than re-litigated context, and reusing previously accepted answers instead of resending them.

Implications

• Context/token efficiency: names four concrete, reusable techniques (clean tasks, evidence selection, settled-steps-as-code, accepted-answer reuse) for bounding the token growth curve of long agent sessions — directly applicable to any workflow that resumes or continues state across turns.
• Agentic engineering patterns: “settled steps as code” is notable as a pattern — pushing work that no longer needs to be re-reasoned about out of the LLM context entirely and into deterministic code, rather than just compressing the prompt.