Here's a number that should make AI cheaper by now: the cost of running a model with GPT-3.5-level capability has fallen roughly 280-fold in two years, according to Stanford HAI's AI Index. A task that cost a dollar in late 2022 costs well under a cent today. And yet, enterprise AI spending isn't falling. It's climbing fast enough that CIOs are starting to budget for it the way they'd budget for a volatile commodity, not a predictable software line item.
Quick answer: AI inference costs have fallen dramatically, by roughly 280x over two years according to Stanford HAI, yet total AI spending keeps rising because usage is growing even faster than prices are falling. Economists call this the Jevons Paradox: when something becomes more efficient to use, people don't use the same amount for less money, they find more uses for it, and total consumption rises anyway. For developers, the practical driver is that agentic workflows, where an AI plans, calls tools, and checks its own work, can burn 5 to 1,000 times more tokens than a single simple query.
The Number That Doesn't Add Up, Until You Look Closer
The price collapse itself isn't in dispute. Multiple sources converge on the same broad pattern, even if the exact multiplier varies depending on which model generation and time window they're measuring. Stanford HAI's widely cited figure is 280x over two years for GPT-3.5-equivalent performance. A separate analysis from Epoch AI describes a roughly 10x-per-year decline for equivalent-capability models, which compounds to similar territory over a two-year span. Other outlets have reported figures as high as 600x depending on which specific model comparison they use. For a fixed level of capability, AI has gotten dramatically, almost absurdly, cheaper to run. The exact multiplier matters less than that direction.
Gartner's own worldwide AI spending forecast puts total 2026 enterprise AI spend in the trillions, still climbing year over year. One analysis tracking a single company's inference bill found it tripled even as its per-token price fell by roughly that same order of magnitude. There's nothing contradictory buried in that data. It's a predictable economic pattern with a name older than computing itself.
This Already Happened Once, With Coal
In 1865, economist William Stanley Jevons noticed something counterintuitive about steam engines. As engineers made them more fuel-efficient, Britain's coal consumption didn't drop. It rose, because cheaper-to-run engines got installed in more factories, more ships, and more railways than before. Efficiency didn't shrink demand. It expanded the set of things worth doing with the resource, and total consumption climbed.
Apply that same logic to AI tokens and the pattern repeats almost exactly. When a task cost a dollar in API fees, a lot of exploratory or marginal uses simply weren't worth trying. When that same task costs a fraction of a cent, it becomes worth running speculatively, running multiple times to compare outputs, or embedding into a workflow that checks its own answer before responding.
Where do the extra tokens actually go?
That's the half that matters most if you're building with AI tools, and I've covered it on my blog:
- Why one agentic task can burn 5x to 1,000x more tokens than a simple query
- What a 2026 CIO survey says enterprises will do when token prices fall (95% said they'd spend more)
- Why inference is now 60-85% of enterprise AI budgets
- A practical way to scope agent tasks so your own API bill stays under control, plus an FAQ
👉 Read the full article: Why Are AI Bills Still Rising? The Inference Cost Paradox Explained
Originally published on NextGentech-official.
Top comments (0)