Cheaper Tokens, Bigger Bills: The Jevons Paradox of AI
AI token prices fell about 280x in two years, yet compute spend keeps setting records. Why cheaper intelligence means bigger bills, seen from my own usage.
Figures checked on 8 November 2026.
Two numbers, a few years apart.
When GPT-4 launched in March 2023, OpenAI charged $30 per million input tokens and $60 per million output tokens. Since then the price of a fixed level of capability has collapsed. Stanford's 2025 AI Index estimated that the cost of querying a model scoring at GPT-3.5 level on the MMLU benchmark fell from about $20 per million tokens in November 2022 to about $0.07 by October 2024, roughly a 280-fold drop in under two years.
Now the other number. After their October 2026 earnings, the largest cloud and AI companies were guiding to well over $700 billion of combined capital spending for 2026, most of it on AI data centres, and several said 2027 would be higher.
The price of a unit of intelligence fell by two orders of magnitude, and the total spend on intelligence went up. That isn't a contradiction. It's one of the oldest results in resource economics.
Jevons, in one paragraph
In 1865, William Stanley Jevons noticed that more efficient steam engines did not reduce Britain's coal consumption. Consumption kept rising. Cheaper steam power made steam worth using in places it had never made sense before, and the new uses swamped the savings from the old ones.
Tokens behave the same way. Every price cut moves a tier of previously uneconomic ideas into the "obviously worth doing" column, so the total bill rises even as each call gets cheaper.
The paradox only holds when demand is highly elastic and far from saturated. That describes AI inference in 2026. It won't describe it forever, and I'll come back to that.
I am the paradox
I don't need industry data to see this. My own usage shows it.
Between April and October 2026 I moved my engineering work onto Claude Code agents. I wrote it up in detail in my six-month case study. In the eleven weeks from mid-July I logged about 400 hours of agent-assisted work over 83 active days: 513 sessions and 978 delegated sub-agent runs, often two or more sessions in parallel.
Before mid-2026 my agent use was much lighter. As the tools got cheaper and better, I didn't use AI less. I used it for far more things: searching codebases, writing tests, reviewing architecture, running scheduled morning briefs and weekly health checks. Each one is cheap on its own. Together they're a workload that didn't exist before.
The same pattern shows up in my cost data. A two-week audit found 86% of my sub-agent work running on mid-tier or small models, because most agent work is reading and searching. A command proxy cut token use by another 34%. Savings like these don't shrink usage. They make more of it worth doing. That's Jevons from the inside. I wrote up what actually drives that bill separately.
Why prices fell
Four things happened at once, and none of them is magic.
Distillation and smaller models. Once you have an expensive frontier model, you can use it to train much smaller models that keep most of the useful behaviour on routine tasks. Last year's frontier becomes this year's mid-tier, at a fraction of the cost.
Open-weight competition. Capable open-weight models from Meta, Mistral, Alibaba's Qwen team and DeepSeek mean there is always a self-hostable alternative. That disciplines closed providers' prices whether or not you ever deploy one.
Inference hardware. NVIDIA still dominates, but Google's TPUs, Amazon's Trainium and Microsoft's Maia chips are all in production use for inference, which is where volume is growing fastest.
Better serving software. Continuous batching, paged attention, speculative decoding, cache reuse and quantised weights. Unglamorous engineering that multiplied tokens per GPU-hour on the same hardware.
Why volume exploded anyway
At Google I/O in May 2026, Sundar Pichai said Google was processing more than 3.2 quadrillion tokens a month, about seven times the figure a year earlier.
Three forces drive that.
Agents are token-hungry. A chat turn is one request and one response. An agent task is a loop, and every pass re-sends the accumulated context: history, tool outputs, file contents. A 20-step agent run costs far more than 20 chat turns, because context compounds. Most of the bill is input the agent reads, not code it writes.
Every product now has an AI feature. When a call cost a few cents, adding AI needed a business case. At today's prices it's a default: summarise this, draft that, explain this chart. Each feature is cheap. There are now a great many of them.
Reasoning models trade tokens for quality. Extended thinking spends extra tokens to get a better answer. Product teams have found they like turning that dial up.
Who captures the value
Follow the money and much of it lands with the companies that make and host the machines. In its quarter ended July 2026, NVIDIA reported revenue of $96.2 billion, of which $89.0 billion was data centre, more than double a year earlier.
The cloud providers turn the same demand into cloud growth, and they are putting their own chips into the inference tier to keep more of the margin.
Model labs have a harder business. Training costs scale with ambition, inference prices are set by whoever prices most aggressively, and open weights cap what anyone can charge for commodity capability. Selling tokens is a tougher business than selling the machines that make them.
Wrapper startups get squeezed. If your margin is the spread between what a model costs you and what you charge, falling prices sound good. But customers expect the cut passed on, while your own tokens per user rise as you add agent features.
Enterprises without cost controls get surprised. Teams budget AI per seat, then ship an agent that consumes millions of tokens a day. The unit price fell, so nobody put a meter on it. In May 2026, Fortune reported that Uber had used up its 2026 AI coding tools budget in four months. Cheap tokens are exactly why the volume needs watching.
The counterargument
Jevons effects end. The mechanism is saturation. Demand is elastic only while there are unserved uses. Once every document is summarised and every ticket triaged, further price cuts just reduce the bill. On-device inference, where small models run on a phone or laptop at no cost to the provider, could also flatten cloud demand.
I find that plausible, and I don't see it in the data yet. The sign to watch for is token volume growth slowing while prices are still falling. As of late 2026, volume and spending are both still climbing.
What to do about it
If you're building: design for volume, not price. Measure tokens per task, not per call, because the agent loop is where budgets go. Route by task: a cheaper model for reading, searching and extraction, the expensive one only where judgement matters. Assume your per-user usage will rise faster than unit prices fall.
If you're buying: negotiate on committed volume, avoid long lock-in at today's rates, and insist on per-team usage reporting before you scale any agent rollout.
If you're investing: I'm not a licensed financial adviser and this isn't investment advice. The frame I'd use is that in a Jevons regime, the scarce inputs are compute, power and chips, not tokens. That says nothing about whether any company's valuation is sensible, and the whole thesis depends on demand continuing to grow.
The cheapest thing in AI right now is the intelligence. Everything around it, the chips, the data centres, the electricity, is getting more expensive and more contested. That is what a Jevons paradox looks like from the inside.
Related Reading
Enjoying this article?
Get posts like this in your inbox. No spam, unsubscribe anytime.
