Two sentences, both true, held side by side:
Inference for constant capability is falling roughly ten times a year, and has collapsed by orders of magnitude since 2023. And almost every organization touching this technology is watching its total AI bill grow.
The instinct is to assume one sentence cancels the other, and the instinct is wrong. The second sentence is caused by the first. Understanding that mechanism is the difference between a team that rides the cost curve and a team that is ridden by it.
The falling half
The cleanest statement of the decline came from a16z in November 2024, which called it LLMflation: for models of equivalent performance, inference cost drops about ten times per year. What cost tens of dollars per million tokens a few years ago now costs cents for the same measured quality. Epoch AI's analysis agrees on the direction and adds honesty about the spread: prices fell by a median of about fifty times across benchmarks, but unevenly, with some tasks falling much faster than others.
The precise multiple matters less than the shape. The price of a unit of intelligence behaves like nothing else in a typical budget. It is a falling object, and it has been falling for years.
The growing half
While the unit price falls, the total bill grows, and the causes are structural:
- Volume eats the discount. The same task done ten thousand times instead of ten times leaves the bill larger even after the unit price has fallen by half. A voice call that cost seven dollars and now costs forty cents does not save the sales floor money. It changes the number of calls.
- Ambition climbs to the new price. Tasks that were absurd at last year's prices become obvious at this year's, and each generation of models makes the next tier of work feel one step out of reach, priced at one step away.
- Parallelism multiplies usage. A single chat became an agent, the agent became a lead with subagents, and the same user now consumes on the order of fifteen times the tokens per task. Anthropic's own numbers, 2025.
This pattern is old enough to have a name. When coal got cheaper, England did not use less coal. It found new things coal could do, and used more coal than ever. Cheaper units expand consumption faster than the savings arrive. Economists call it Jevons paradox, and AI is its cleanest modern case.
What the paradox changes about planning
Teams that miss the mechanism plan badly in a specific, predictable way:
- They budget by today's prices and assume savings. The savings are real and they are absorbed by volume before they reach the bottom line. Budget the experiments you want to run, not the price per run.
- They celebrate the falling unit price in public and grow the total bill in private, then treat the growth as a surprise. It is not a surprise. It is the mechanism.
- They delay work waiting for prices to fall. With a ten-times-a-year curve, waiting a year is a real strategy for some marginal projects, but the projects that compound, the ones with feedback loops and owned data, reward starting now at a small scale more than starting later at a large one.
The discipline that works is to plan in three lines: what the work costs at today's unit prices, what volume the new ambition implies, and what the verification of all that output costs. The third line is the one teams forget, and it is the one that does not fall with the price of generation.
The line that does not fall
Here is the asymmetry that should organize the whole budget. Generation keeps getting cheaper. Verification does not, because verification is judgment, and judgment is human, trained, and expensive on purpose.
A falling generation price with a flat verification price moves the whole economic center of the operation toward review, sampling, and evidence. Cheap output is only cheap if accepting it does not cost more than producing it. The organizations that internalize this early spend their growing budgets on verification capacity and ownership, and the ones that do not spend them on generating things nobody had the judgment to check.
So the bill grows either way. The choice is what it grows into.
FAQ
If prices fall ten times a year, should we just wait before committing?
For bounded, one-off tasks with no learning attached, waiting is often rational, and the curve is the reason. For anything with compounding value, prompt libraries, eval suites, feedback data, ownership habits, waiting costs more than it saves, because those assets take cycles to build and they get better while prices fall underneath them.
How do we keep the total bill from growing blindly?
Make volume a decision with an owner, the same way the [calling floor became a product](/insights/one-minute-one-hundred-calls). Track cost per outcome rather than cost per token, cap experiments by expected evidence rather than by budget vibes, and review the top three consumers monthly. Most bill growth is three line items nobody is looking at.
Does the falling price change what we should attempt?
It changes it constantly, and that is the opportunity. Keep a list of tasks you rejected eighteen months ago as too expensive, and re-price them every quarter at current unit costs. Several will have crossed the line. The falling curve is not just a discount on what you already do. It is a door opening on what you never priced in.
Is there a floor where the decline stops?
Energy, hardware, and inference efficiency all still have room, and every year the floor prediction has been wrong. Plan on the decline continuing unevenly, re-check the load-bearing numbers quarterly with their dates, and hold your strategy loosely enough that a ten-times change in either direction does not break it.
Related reading
- The Swarm Is a Different Animal
- Date-Stamp the Numbers
- Speed Is Not the Multiplier
- The Review Bottleneck Is the Real Cost Center in AI Teams
- One Minute, One Hundred Calls
- The New Venture Stack: Small Teams and Agent Fleets
If your AI bill is growing faster than your clarity about why, book a discovery call.