Cheaper LLM Tokens Led to Bigger AI Bills as Jevons Paradox Takes Hold
Key Takeaways
- ▸Token price drops of ~80% paradoxically increased total AI spending rather than reducing it, following the Jevons Paradox pattern
- ▸Agentic workloads multiply token consumption far beyond simple chat interactions; single tasks consume 1–3.5M tokens and power users can spend $1,800+ monthly
- ▸Output token pricing (4–10× input cost) is the primary cost driver for agentic systems, making output volume the critical budget variable
Summary
As LLM token prices dropped roughly 80% between 2025 and 2026, organizations expected to see AI costs decrease. Instead, the opposite happened: cheaper per-token pricing enabled engineers to use AI more aggressively and frequently, causing overall spending to surge rather than decline. This phenomenon mirrors the historical Jevons Paradox, where efficiency improvements in steam engines led to increased coal consumption rather than conservation. The effect is most pronounced with agentic workloads, where tasks can generate 1–3.5 million tokens, with heavy users reaching $1,800 monthly—exceeding corporate caps like Uber's newly implemented $1,500-per-month hard cap per employee.
The economics reveal a stark pricing asymmetry: output tokens cost 4–10× more than input tokens across major models, and agentic systems dramatically increase output volume. Using Claude Opus 4.8 as an example, a single agentic coding session with reasonable context costs $2.25 per turn; at 40 daily turns across 20 working days, one engineer hits $1,800 monthly. The cost distribution follows a power law, with 63% of organizations now implementing spending controls. Some companies like Microsoft have cancelled employee AI licenses after discovering individual engineers consuming $2,000+ monthly.
- Organizations are implementing hard spending caps to control costs, yet power agentic users—who generate the most business value—face the steepest restrictions
Editorial Opinion
The Jevons Paradox hitting AI spending represents an important inflection point for the industry: cheaper tools don't automatically translate to cheaper operations when usage patterns fundamentally shift. The concentration of both high value and high cost in the 'power agentic user' tail creates a real governance challenge—blunt per-tool caps may inadvertently restrict the engineers who generate the most business value. Organizations need smarter consumption-based governance models that align incentives with outcomes, not just price-based controls.
