Token prices keep falling, and AI bills keep rising: elasticity decides which wins

In 1865 the economist William Stanley Jevons published "The Coal Question," which noted that improvements in the efficiency of steam engines had not reduced Britain's consumption of coal. They had raised it. Cheaper power per ton of coal made engines worth using in more places, and the extra uses outweighed the saving on each. Satya Nadella invoked the observation in January 2025, when the release of DeepSeek's R1 model made investors wonder whether cheaper AI would reduce the need for data centers. "Jevons paradox strikes again," he wrote, arguing that as AI becomes more efficient and accessible, its use would grow. The remark was convenient for a company that sells AI services, and it should be read with that in mind. The pattern it points to is real, though, and the numbers on tokens show it.

Google reported in May 2024 that it processed 9.7 trillion tokens a month across its products. At its developer conference a year later the figure was 480 trillion, about fifty times higher. By October 2025 Sundar Pichai cited 1.3 quadrillion, and in May 2026 he said 3.2 quadrillion, nearly seven times the 2025 figure. Critics note that these totals measure the work done by Google's infrastructure and not the number of people served, and that the rate of increase slowed during the summer of 2025. Even with those cautions, the direction is clear. Over the same stretch, the price of reaching a given level of capability fell by tens of times per year according to Epoch AI's estimates. Prices fell steeply and volume rose faster.

For anyone who has to budget for AI, the lesson is that the unit price is the less important number. The bill is price times quantity, and the quantity responds to price, to capability and to design. A plan built from a price list and a flat usage forecast is likely to be wrong, and wrong in the unwelcome direction.

The arithmetic of elasticity

Economists describe the response of quantity to price as elasticity. If demand has a constant elasticity ε, a price falling by a factor f raises quantity by f to the power ε, and total spend changes by f to the power ε minus 1. When ε is above 1, a price cut raises spending. When it is below 1, spending falls. At exactly 1, spending is unchanged.

An example with round numbers. A feature handles 1 million requests a month at $0.02 each, for $20,000. A new model or a router cuts the cost per request tenfold to $0.002. If elasticity is 0.5, volume rises to about 3.2 million, and spend falls to $6,300. At 0.7, volume reaches 5 million and spend halves to $10,000. Unit elasticity takes volume to 10 million and leaves spend at $20,000. A value of 1.5 sends volume to 31.6 million and spend to $63,000, more than triple, while 2.0 would bring 100 million requests and a $200,000 bill. The same price cut delivers savings in one scenario and a tenfold increase in the bill in another.

Elasticity is a property of the use case. It tends to be high where the price is what keeps a task from being done, and the supply of tasks is large: classifying every incoming document instead of a sample, summarizing every meeting, checking every transaction, drafting a personalized message for each customer. Each of these has a price below which it becomes worth doing for all items, and above which only for a few. It tends to be low where volume is capped by something else, such as the number of customers a team can serve or the number of regulatory filings due.

Four ways consumption outruns price

Several mechanisms push volume up faster than price falls, and they are worth separating.

The first is new use cases crossing a viability threshold. A task that costs $0.50 per item is worth doing for high-value items and not for the rest. At $0.005 it is worth doing for everything. The market for the activity does not grow by a factor of a hundred; the set of items for which it makes sense does.

The second is tokens per task rising. Reasoning models generate long internal chains before answering, and those tokens are billed as output. Agents loop through tools and revise. Anthropic's account of its research system reports that agents use about four times the tokens of a chat, and multi-agent systems about fifteen times. A price per token that falls by half can be swamped by a task that uses ten times as many tokens.

The third is longer context. Retrieval, memory, system prompts and conversation history are resent on every call. A request that once carried a couple of thousand tokens can carry tens of thousands, because the models can use them and the product team can afford them.

The fourth is ambient use. Once a capability is cheap, it is added everywhere: every email gets a summary, every ticket a suggested reply, every document a tag. Each is a small call, and the number of them is large.

A bill that rises while prices fall

Here is a made-up support assistant followed over three years. In year one it handles half of the company's 2 million annual tickets, with a single call of about 4,000 tokens per ticket at a blended $5 per million tokens, which is $0.02 a ticket and $20,000 for 1 million tickets. In year two the team moves to a cheaper model at $2 per million tokens, a 60% cut. It also adds retrieval over the knowledge base and longer histories, so each ticket now uses 12,000 tokens, and extends coverage to all 2 million tickets. The bill is 2 million × 12,000 × $2 per million, or $48,000. In year three the price falls to $1 per million. The team adds an agent that looks up orders and drafts refunds, using 40,000 tokens per ticket, and opens the email and chat channels, which brings the volume to 3 million tickets. The bill is 3 million × 40,000 × $1 per million, or $120,000.

Year	Price per million tokens	Tokens per ticket	Tickets	Annual bill
1	$5	4,000	1 million	$20,000
2	$2	12,000	2 million	$48,000
3	$1	40,000	3 million	$120,000

The price per token fell five times, and the bill rose six times. The factors multiply: tokens per ticket grew tenfold, tickets tripled, and the price fell to a fifth, giving 10 × 3 × 0.2 = 6. Cost per ticket doubled from $0.02 to $0.04, and whether that is good news depends on what the ticket gets. If the assistant now resolves 70% of cases end to end where before it drafted suggestions that agents rewrote, the economics improved while the bill grew. The point of decomposing the bill into price, intensity and volume is to see which factor moved and to judge whether each was a decision someone made.

The ceiling moves too

The price decline Epoch measures applies to a fixed level of capability. A company that wants the best available model faces a different series, because new tiers appear at the top. Anthropic's price list in October 2026 shows the earlier flagship Opus 4.1 at $15 per million input tokens and $75 per million output tokens, now retired except on certain cloud platforms, the current Opus 5.5 at $4 and $20, and a higher tier above it, Fable 5.1, at $10 and $50. A buyer who held the same capability level paid far less over that period. A buyer who always bought the top of the range saw the price of the top rise. Which series applies to a given product depends on whether it needs the frontier or can use what the frontier did a year ago, and the answer is a design choice that belongs in the budget.

Industry totals tell the same story from the supplier side. Menlo Ventures estimated, in a survey reported in mid-2025, that enterprise spending on LLM APIs rose from about $3.5 billion in November 2024 to $8.4 billion by mid-2025, in a period when prices of comparable models fell. The estimate comes from a survey and carries the usual limits, and it is consistent with volume growing faster than price declines.

Where the paradox does not apply

Jevons' observation is not a law of nature. Energy economists who study the rebound effect usually find that efficiency gains in a given use, such as home heating or driving, are partly offset by increased use and rarely more than offset. Full backfire, where total consumption rises, is a special case that occurs when demand is highly responsive and the resource is a general-purpose input. AI inference appears to fit the pattern, since it is general-purpose and the demand for cognitive tasks looks highly elastic so far, but individual applications will differ.

Inside a company the dynamics are different again. When employees use a tool under a flat license, they see no price at all, and demand is limited by attention and workflow, not cost. Usage then grows with capability and adoption. The vendor's per-seat price is flat, but the vendor's cost per seat is not, as the heaviest users of agent features show. That produces a pricing problem for sellers and a metering problem for buyers, and both lead to the same remedy.

What this means for planning

Several practices follow from the arithmetic.

Budget from workflows and volumes. For each use case, estimate the number of units of work per month, the tokens per unit under the current design and the price per token, and treat all three as moving. Then run the scenarios: volume up 3 times, up 10 times, tokens per unit up 4 times because someone adds an agent. A budget that cannot absorb the second scenario is not a budget.

Measure cost where it is incurred. Tag every call with the feature, the customer and, where possible, the outcome it served. Cost per resolved ticket, per processed claim or per drafted document is a number a business owner can reason about. Cost per million tokens is not. Set alerts on cost per unit as well as on total spend, because a rise in either has a different cause.

Put limits where the use is open-ended. Rate limits, per-user quotas, caps on agent steps and spending ceilings per task turn an unbounded liability into a bounded one. Design the fallback: when a ceiling is reached, the system degrades to a cheaper model or a shorter answer instead of failing.

Spend the savings on purpose. A price drop can be taken as a saving, or reinvested in capability, such as a stronger model for the same cost or more coverage of the workload. Make that choice explicitly. Companies that pocket every saving may find their competitors have spent theirs on a better product.

Negotiate contracts for a falling market. Committed-spend discounts made sense when prices were stable. When the price of a given capability can fall by half within a year, a two-year commitment at today's rate may lock in an above-market price. Prefer shorter terms, price-protection clauses that follow the vendor's list prices downward, and flexibility to move volume to a newer model.

AI spend tends to be nobody's job. The engineering team sees a line in the cloud invoice, finance sees a vendor, and the product owners see features. Give the bill an owner, usually a product or platform lead with a finance partner, and a monthly ritual: review cost per unit by feature, compare it with the forecast, and explain every variance larger than a set threshold. Include anomaly detection, because the most expensive failures are accidents. An agent stuck in a retry loop, a prompt that accidentally includes a whole database table, or a batch job launched twice can burn a month's budget in an afternoon, and a hard spending ceiling per task and per day turns such an accident into an incident report instead of a surprise on the invoice.

Estimating your own elasticity

Elasticity can be estimated from your own data. Where a product has changed its effective cost per request, through a model switch, a router or a caching change, compare volume before and after, controlling for seasonality and for any other changes. A rough log-log relationship over a few periods gives a first estimate. Where no such history exists, run a pilot with different default settings across teams or cohorts. The result is informative even if imprecise, because the decision it supports is coarse: whether to plan for a bill that falls, stays flat or grows as prices drop.

A planning habit

Add one line to every AI business case: what happens to total cost if the unit cost falls by 10 times? If the answer is "the bill falls," state the assumption about volume that makes it true. If the answer is "the bill rises," say whether the value created rises faster. A case that has this line demonstrates that its authors have thought about the customer's behavior as well as the vendor's price list.

Sources

1. Jevons, W. S. (1865). The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. Macmillan.
2. Nadella, S. (27 January 2025). Post on Jevons paradox following the release of DeepSeek R1, as reported by Fortune, GeekWire and NPR Planet Money.
3. Google I/O 2025 and I/O 2026, and Google Cloud event (October 2025): reported monthly token volumes of 9.7 trillion (May 2024), 480 trillion (May 2025), 1.3 quadrillion (October 2025) and 3.2 quadrillion (May 2026), as reported by the specialized press, with caveats from The Decoder and others on interpreting the metric.
4. Epoch AI. LLM inference price trends. epoch.ai/data-insights/llm-inference-price-trends.
5. Anthropic (June 2025). How we built our multi-agent research system. anthropic.com/engineering/multi-agent-research-system.
6. Sorrell, S. (2007). The Rebound Effect: an assessment of the evidence for economy-wide energy savings from improved energy efficiency. UK Energy Research Centre.
7. Illustrative elasticity scenarios: one million requests a month at $0.02, a tenfold price cut, constant-elasticity demand, as stated in the text.
8. Menlo Ventures (July 2025). 2025 Mid-Year LLM Market Update, as reported by BigDATAwire and TechCrunch.
9. Anthropic. Pricing, Claude developer documentation (list prices as of October 2026). platform.claude.com/docs/en/about-claude/pricing.
10. Illustrative support-assistant bill: prices, tokens per ticket and volumes stated in the text.