← sources

AMA: I burned 500M tokens optimizing token usage.

source: reddit (reddit:ClaudeCode)published 2026-09-08 17:37ingested 2026-09-09 02:41

retained text

I spent most of the weekend + over 500M tokens auditing and optimizing my token usage. I mostly work in the Claude Code app and have Claude dispatch to CODEX for code reviews; sometimes work directly in CODEX app. **AMA**

**What I found:**  
\- I waste a lot of tokens by returning to long-running sessions after my 1-hr cache TTL  
\- Fable is most effective in long-running sessions that require large context reads and little coding or writing. Mostly synthesis and decisions that require large context windows.  
\- My agents were not resuming subagent threads for multi-round turns (I thought they were). Cache TTL is only 5-min, so I’m uncertain how much this matters

**What I’m doing and testing:**  
\- Built comprehensive token cost tracking for all providers, repos, and tasks with a dashboard I can monitor.  
\- Modifying agent delegation rules based on an audit of costs and usage over the last several months  
\- Surfacing live token cost stats to agents to help them dynamically adjust delegation strategy and pace to my weekly budget  
\- Agent orchestration: Working with multiple Opus instances per repo for development and having Opus delegate generously to Sonnet. Also encouraging Sonnet to delegate to Haiku  
\- Working with 1 Fable instance per repo who oversees work of all Opus sessions in the repo. The goal is to keep Fable cache warm as much as possible within the 1-hr TTL and minimize writes. Each Opus session can automatically message the Fable session. Opus escalates to Fable. Fable escalates to me.  
\- Created a hook for Opus models to send me a compaction instruction after a long-running task completes or at a good stopping point after a context threshold has been hit. I still have to run it manually but it saves time while (hopefully) providing cheap compaction while cache is warm  
\- Use resume more often on subagent delegation to improve cache-use and reduce reorientation on each dispatch  
\- I’ll use the dashboard to track token use across all applications, providers, projects, and tasks to help adjust this strategy over time

What are some of the highest impact and non-obvious things you’ve found?

document fetched 2026-09-09 10:30 · canonical https://reddit.com/r/ClaudeCode/comments/1wav7dw/ama_i_burned_500m_tokens_optimizing_token_usage