Why is my AI bill going up when token prices keep dropping? If you have asked that in a budget review this year, you are not misreading the invoice. Consumption sets the number on the bill: how many tokens each task burns, and how many tasks you run.
The gap between a falling price and a rising bill is the story. Token economics is the study of how tokens are priced, consumed, metered and billed, from the pricing page to the invoice. A token economics overview frames everything that follows. It is the difference between explaining a rising AI bill and defending it with a unit price.
The phenomenon has a name: the inference paradox. What follows unpacks it and hands you the vocabulary to explain a rising AI bill.
Why are token prices falling while enterprise AI costs keep rising?
A falling cost per token is a price signal. Your bill is a consumption signal, set by how many tokens you consume, and that volume is growing faster than the price is falling.
The price falls because open-weight models from DeepSeek and Qwen keep compressing the per-token rate. That is token deflation: a sustained fall in what a token costs, driven by open-weight models, competition and efficiency gains. Silicon Data’s LLM Token Expenditure Index hit a record 97 cents per million tokens on 31 August 2026, down more than half from its summer peak. The index measures what the market actually pays: an expenditure-weighted price across 400-plus models and 20-plus data sources, rather than provider list prices. You can follow it on Bloomberg under SDLLMTK.
Yet enterprise inference still cost $1.16 to $1.18 per million tokens in early August, according to Jefferies. The blended market rate and enterprise list prices diverge, and that gap is where the extra spend hides.
Meanwhile the invoice climbs. Uber exhausted its full-year AI budget by April, and enterprise AI bills have tripled even as per-token prices fell 98%. Inference is now the majority of AI compute cost, roughly two-thirds in 2026. Training is a one-off outlay. Inference is the fuel bill, paid every day. And the cheaper the fuel, the bigger the bill. That combination has a name.
What is the “inference paradox” in enterprise AI economics?
It is the coexistence of falling cost per token with rising total inference spend. Cheaper unit prices, bigger bills. Gartner defines the paradox as better unit economics escalating overall AI cost without a clear path to commensurate value.
The paradox collapses the moment you stop confusing the two quantities. Cost per token is a unit price. Cost per unit of work is tokens per task multiplied by cost per token. Only one side of that equation is falling, so the product still rises.
Gartner forecasts agentic workflow inference cost to rise more than fivefold through 2028, even as token costs fall by roughly 95% by 2030. Those two numbers, fivefold up on spend and 95% down on token cost, are the paradox in miniature, and the vocabulary to use the next time the bill is questioned. Measuring the true cost per task turns the argument into a number.
Why does Jevons paradox explain rising token consumption in AI?
In 1865, William Stanley Jevons observed that when the steam engine made coal cheaper to use, Britain consumed more coal, not less. Efficiency lowered the cost of a unit of work, expanded demand for that work, and raised total consumption. That is Jevons paradox, the special case where demand elasticity pushes total use up rather than down.
Cheaper tokens make always-on agents, retrieval-heavy patterns, retries and long-context prompts economically viable for the first time. Each of those patterns consumes many more tokens than the chatbot-era baseline. A request that used to be one prompt and one answer is now longer contexts, multi-step reasoning traces, chain-of-thought, tool-calling loops and agentic retries. Every step re-sends context and re-runs the model.
As Bain’s analysts put it, the models get cheaper, the usage gets heavier, the bill stays stubbornly high. Token costs halved over a year while tokens consumed grew 450%. Gartner’s fivefold forecast is evidence the pattern is structural, not a fad.
The model layer is now the cheap part, which is why the levers live in the control plane: model routing in production sends simple requests to cheaper models, and semantic caching serves repeat queries without re-invoking a model. The decisions that set your bill now live one layer up.
Falling token prices and rising AI bills are two sides of the same mechanism. The model layer has gotten cheaper, so the savings and the control have moved into the control plane: routing, caching and governance. Track the consumption signal, and you will see what your budget actually moves with: cost per unit of work, tokens per task multiplied by cost per token. The wider inference cost landscape maps the levers, from measuring true cost per task to deciding whether to build or buy the cost layer.
Frequently Asked Questions
What is the LLM Token Expenditure Index and what does it actually measure?
It is Silicon Data’s index of the realised, expenditure-weighted price per million tokens across more than 400 models and 20-plus data sources. Crucially, it tracks what organisations actually pay, not list price. The index hit a record low of 97 cents per million tokens on 31 August 2026, down more than half from its summer peak, and you can follow it on Bloomberg under the ticker SDLLMTK.
Why is inference now the majority of total AI compute cost?
The shift is structural, not seasonal. Training is a one-off outlay that ends when a model ships, while inference is a continuous, variable expense that compounds with every request and every user. As more workloads move into production, the sustained cost of serving tokens simply dwarfs the episodic cost of building the model, which is why inference now dominates the compute bill.
What drives the rise in tokens consumed per request?
Several things stack together: longer contexts, multi-step reasoning traces, chain-of-thought prompts, tool-calling loops and agentic retries. Each iteration re-sends context and re-runs the model, so a single user-visible task can consume many times the tokens of a chatbot-era exchange. The per-request count keeps climbing even when the price of each individual token keeps falling.
Where can I track leading LLM token price indices and benchmarks?
The most cited is Silicon Data’s LLM Token Expenditure Index, available on the Bloomberg terminal under SDLLMTK, alongside Silicon Data’s published benchmarks. Pair it with provider pricing pages and enterprise contract rates, because blended market indices and enterprise list prices diverge. Tracking both tells you whether your own effective rate is keeping pace with the market.
Does a falling token price actually mean cheaper AI overall?
No, and that is the whole point of the paradox. A falling price per token is a price signal, not a bill signal. Your total spend is set by consumption, which is tokens per task multiplied by the number of tasks. Because tokens per task is growing faster than the price is falling, the product still rises, and so does the invoice.
Is the inference paradox just short-term inflation that will correct itself?
Not really. The mechanism behind it, Jevons paradox, is structural rather than cyclical. Cheaper tokens expand the set of economically viable uses, from always-on agents to retrieval-heavy patterns, and those new uses consume more tokens than the workloads they replace. Gartner expects agentic workflow inference cost to rise more than fivefold through 2028 even as token cost declines by roughly 95% by 2030.
How do I calculate cost per unit of work for my own AI workloads?
Start by measuring tokens per task, not tokens per call: count every prompt, completion, retry and tool call that a single completed task requires. Multiply that by your effective cost per token, then divide by the number of completed tasks to get a cost per unit of work. That figure, not the list price, is the number your budget actually tracks.
What happens if I keep budgeting for AI on falling per-token prices?
You will keep getting blindsided. Organisations that assumed falling prices would flatten their spend have already hit three times their annual AI budgets by June 2026, and Uber exhausted a full-year AI budget by April 2026. Token spend climbed roughly 497% between January 2025 and April 2026. Budget for consumption growth, and treat price decline as a partial offset at best.
Does training a model still matter if inference dominates the cost?
Yes, but it matters differently. Training is a capital-style, one-off investment that you amortise; inference is an operating expense that repeats and compounds every month. Inference is now the majority of total AI compute cost, so the leverage for controlling spend has moved from the training run to the serving path: how you route, cache and govern live traffic.
Will open-weight models like DeepSeek lower my enterprise AI bill?
Not automatically, and the numbers show why. Open-weight models and provider price cuts have compressed the blended market rate to a record 97 cents per million tokens, yet enterprise inference still costs around $1.16 to $1.18 per million tokens. The market index and enterprise rates diverge, so your realised saving depends on where and how you deploy, not on the headline price.
Can caching and routing offset rising token consumption?
They blunt it rather than reverse it. Model routing sends simpler requests to cheaper models, and semantic caching serves repeat queries without re-invoking the model. Both cut tokens consumed per task, which is the side of the equation you can actually control. They will not outrun Jevons-style demand growth on their own, but they shift the savings into the control plane where you can manage them.
What does token deflation mean?
Token deflation is the sustained fall in the price of a token over time, driven by open-weight models, provider competition and efficiency gains. It is real and measurable: the market rate fell more than half from its summer peak to 97 cents per million tokens by 31 August 2026. The trap is treating that deflation as though it guarantees a smaller bill, when expenditure growth can easily outpace it.
