McKinsey Says Token Prices Plummeted. AI Bills Keep Climbing. Jevons Called This One In 1865.
McKinsey experts report that GPT-4 launched in 2023 with its 8K-context API priced at $60 per million output tokens, yet models performing at roughly the same level on established benchmarks now cost a fraction of that amount. Despite collapsing per-token costs, total AI spending continues to rise. McKinsey's Catlin, also a director of the McKinsey Global Institute, notes there is no single cost lever and identifies three areas where companies can manage AI spending, starting with visibility into which use cases and business units consume resources.
This is a textbook demonstration of Jevons Paradox. The mechanism is that increased efficiency in resource consumption drives up total demand enough to offset the per-unit savings. As the cost of producing a unit of intelligence collapses, consumption explodes. The lesson for anyone managing AI spending is that cost per token is a vanity metric. Track which use cases generate value instead of obsessing over unit price.
McKinsey, with Catlin, a director of the McKinsey Global Institute, outlining the spending dynamics and cost management framework.
- Sign up for an AI tool that shows your usage metrics, such as OpenAI's API dashboard or any platform with a usage breakdown. Note your total spend over the past month.
- Categorize what you actually used AI for across that period. Label each session as one of three types: work, learning, or entertainment. Tally the approximate cost per category.
- Identify the category that consumed the most budget and ask whether it produced proportional value. If not, cut that usage pattern next week and watch whether your total spend drops. The point is not cheaper tokens but smarter consumption.