You’re Loading 66,000 Tokens of Plugins Before You Even Type. That’s Why Your Limit Disappears.

Source: Nate’s Newsletter Date: 2026-04-02 URL: https://natesnewsletter.substack.com/p/your-claude-sessions-cost-10x-what

Summary

Most Claude users waste 5-20x more tokens than necessary through habits carried over from ChatGPT, with the most expensive single pattern being naively loaded plugin/tool contexts. 66,000 tokens of plugins loaded before a single user message means the context window is functionally depleted before work begins. A sophisticated multi-conversation production pipeline costs less than a quarter per user; typical users spend more asking basic questions. The problem is habit, not infrastructure.

Implications

Agent-product positioning / enterprise adoption thread. Token waste at this scale is a user-education problem that looks like an infrastructure problem, which makes it expensive to diagnose and easy to blame on Anthropic. For teams building production systems, this is an architecture argument: context budget management (one of the 12 infrastructure primitives) needs to be a first-class concern, not an afterthought. For Anthropic, users who hit limits due to poor habits churn or complain publicly — a customer success surface that tooling could address.

• Pressures: any agent orchestration layer that doesn’t expose context budget usage to operators is hiding a cost control problem; enterprise buyers will eventually surface this in RFPs.
• Watch: whether Anthropic ships any token-usage visibility or budget management tooling into the core Claude interface.