Cheaper AI Tokens Are Actually Costing Companies More, Not Less

Here's a counterintuitive one: as LLM token prices have dropped dramatically over the past two years, total AI spending at many companies has gone up, not down. Northwood Systems laid out the dynamic in a recent blog post, pointing to the Jevons paradox — the 19th-century economic observation that cheaper resources tend to get consumed in larger quantities, wiping out the savings.

The pattern tracks. When GPT-4-class inference cost $30 per million tokens, teams kept usage tight. Now that comparable models run under $1 per million tokens, the guardrails come off — more agents, longer context windows, more retries, more experimental features nobody's shut down yet. The bill climbs anyway.

It's not a bug in the economics, it's a feature of how organizations actually behave. Cheap compute invites sprawl. For engineering and finance teams trying to forecast AI costs, Northwood's post is a useful reality check: don't budget based on per-token price alone.