Your Claude Code Tokens Are Disappearing

This episode of the Authority Hacker Podcast dives into two major developments affecting AI-powered development tools: the massive leak of Claude Code's source code and the tightening of usage limits by Anthropic. The leak—over 500,000 lines of code, including unreleased features, system prompts, and agent architecture—has shattered Anthropic’s competitive moat, enabling rivals like OpenAI and Cursor to rapidly replicate advanced features such as Kairos (auto-sleep mode), Ultra Plan (cloud-based planning), and undercover mode. While this undermines Anthropic’s pricing power, it benefits users by increasing competition and lowering switching costs. Simultaneously, Anthropic has silently reduced token usage during peak hours, creating a 'peak hour multiplier' that burns tokens faster, affecting U.S.-based users most severely. The hosts offer practical strategies to stretch tokens—using cheaper models like Haiku and Sonnet, optimizing skills and cloud.md files, scheduling tasks during off-peak hours, and minimizing chat history bloat. They also spotlight Cloudflare’s new open-source CMS, Emdash, a WordPress alternative built on Astro with serverless architecture, AI-first design, and built-in security and monetization features. Finally, the episode examines OpenAI’s new self-serve ad platform in ChatGPT, with high CPMs but low CTRs, and speculates on a future split between a monetized B2C ChatGPT and a premium, ad-free professional app. The hosts emphasize that the era of cheap, unlimited AI tokens is ending, and efficiency will soon be critical for small businesses.