Payden & Rygel: AI token costs are falling, potentially deflationary
With the flurry of inflation data released this week, we continue to hear inflation fears related to data center buildouts — everything from higher electricity prices to more expensive hardware. However, as we always say, "it's more complicated!"
The cost of using Al models that output tokens to the end user (consumer) may not be rising but falling. What?! Yes, as with the measurement of goods and services inflation you must adjust for quality improvements. With regard to Al, that means gauging the token cost per unit of a constant or adjusting the level of intelligence.
Turns out, for all models that have the same intelligence level as ChatGPT-4, released in November 2023, the cost to produce a million tokens has fallen over 99.9% from the first announced $60 to only $0.05 in January 2026, not to mention that some of these models are far outperforming ChatGPT-4!
In turn, token deflation means Al usage is getting cheaper, generating better output, and increasing overall efficiency, which, if sustained in the long run, can be a deflationary productivity boost.
