The Jevons Paradox in AI: Why Cheaper Tokens Are Burning Bigger Budgets
As inference costs plummet, AI usage is exploding. The economic principle from 1865 explains why your AI bill keeps going up, not down.
In 1865, William Stanley Jevons published a book about coal. He argued that making steam engines more efficient wouldn’t reduce coal consumption — it would increase it. Cheaper coal meant more engines, more industry, more demand. The Jevons Paradox was born.
160 years later, it’s happening to AI.
The cost collapse
OpenAI’s GPT-4 launched in March 2023 at $36 per million input tokens. By mid-2026, comparable models cost under $2. That’s a 95% drop in three years. DeepSeek, Google, Anthropic, and others have been racing to the bottom on price while climbing on capability.
If you’re running an AI product, this should be great news. Lower costs mean better margins, right?
Not quite
The data tells a different story. Companies are spending more on AI than ever — not less. The reason is simple: when something gets cheaper, people use more of it.
Anthropic reported that token usage on their API grew 15x year-over-year. OpenAI’s enterprise revenue crossed $1B annualized. Individual developers who used to carefully ration API calls now run multi-step agent loops that burn through millions of tokens per task.
When the price drops 95% but usage goes up 1500%, your bill goes up.
Why agents make it worse
The Jevons Paradox is amplified by agentic workflows. A traditional chat completion might use 500 tokens. An agent that thinks, plans, uses tools, and self-corrects can easily consume 50,000+ tokens for the same task.
Each step in an agent loop is cheap — maybe $0.001. But agents run dozens of steps, hundreds of times a day, across thousands of users. The math is brutal:
- Single agent task: ~$0.001
- Per user per day (10 tasks): ~$0.01
- 10,000 users: ~$100/day
- Monthly: ~$3,000
And that’s for a simple agent. Complex reasoning models with extended thinking tokens cost 5-10x more.
The Jevons loop
This creates a feedback loop:
- Models get cheaper → more companies build AI products
- More products → more agent workflows → exponentially more tokens
- More tokens → higher total spend
- Higher spend → demand for cheaper models → back to step 1
The paradox isn’t a bug. It’s the mechanism by which AI infrastructure scales.
What this means
For builders: your AI bill will keep going up. Budget for it. The era of “AI is basically free” is a myth — it’s only free if you don’t use it, and you’ll use it more than you think.
For the industry: the total AI compute market is growing faster than any individual model’s price is falling. Nvidia knows this. Cloud providers know this. The Jevons Paradox is their business model.
Jevons was right about coal. He’s right about tokens. The cheaper it gets, the more we burn.