GPT-5.6 Luna Just Cut Prices 80% Your AI Bill Is Still Going Up, and Here’s the Math

This week in the AI price war, OpenAI slashed GPT-5.6 Luna pricing by 80% and GPT-5.6 Terra by 20%. Meanwhile, Claude Opus 5 appeared on Amazon Bedrock with a pricing range of $5 to $25 per million tokens and a 1M context window. While headlines claim that intelligence is getting cheaper rapidly, engineering teams report an unexpected surge in their AI-related cloud bills.

The reason? The tokens, the only part of the stack experiencing price reductions, are becoming the smallest part of the bill. To illustrate this, the Jevons paradox — which stated that efficiency in steam engines led to increased coal consumption — is applied to tokens. As tokens became 80% cheaper, teams did not reap the savings but instead scaled their workloads that were previously too expensive to run.

This increase in usage resulted in a 2× growth in the bill, not a failure of discipline, but rather the price cut working precisely as intended for the vendor. The AI infrastructure cost structure is shifting, with token spend becoming a minority. New factors such as agent runtime, agent memory and state, observability, and GPU utilization are driving up costs.

FinOps teams need to account for these additional expenses and optimize their cloud resources accordingly. The article concludes by asking if anyone has observed a decrease in total AI spend after a price cut, as the center of gravity of AI costs has moved away from the model and into the supporting infrastructure.