Tokenomics: Why Your $120 AI Plan Is About to Break

If you've been running Claude Code flat out this week and watched your quota vanish, it's not you. The pricing was never built for people who actually use it.

This week on AI Operators, we pull on the loose thread of the Claude Code source leak and end up unpicking the whole economics of AI subscriptions. The leak itself was not a data breach, it was a blueprint of the harness, the deterministic scaffolding Anthropic wraps around the model. Matt walks us through what actually got exposed, Marno explains why your data was never at risk, and Dave connects the dots to the new pricing structure that quietly activates extra usage even when you still have quota.

From there we dig into why specialising your agents, giving them hierarchies and narrow jobs, is the only real defence against runaway token burn. Matt shares why he has fully walked away from Claude Code and rebuilt his workflow on PI. Dave shows the sergeant hierarchy he's building to delegate context instead of dumping it. And Richard reminds us that none of this matters if you cannot answer the only question that counts: what's the real problem you're trying to solve?

We close on the biggest rabbit hole of the week, memory. Matt drops a clean three-part framework (Compilation, Retrieval, Continuity) for how agents should handle context.