Which AI Models Actually Get Cheaper

Two AI features built in the same month can have completely different cost futures. One gets cheaper by half every year or so without anyone touching it. The other costs roughly the same in three years as it does today.

The difference is not how well they were built. It is which tier of model they sit on, and the first systematic economic analysis of token pricing puts numbers on how far apart those two futures are.

The study

Tiered Super-Moore's Law by Mingdeng Du, submitted 30 March 2026, assembles a dataset from OpenRouter API data covering 318 models, Epoch AI records covering 3,237 models, and 62 cross-validated milestone observations spanning 2020 to 2026.

It confirms the headline everyone quotes: roughly a 600-fold decline in token prices. Then it splits that decline by tier, and the aggregate turns out to be hiding the thing you need for a budget.

The projection column is ours, calculated from the paper's half-lives. It assumes the trend continues, which is exactly the assumption you should hold loosely. The paper measures the past; nobody can measure the future.

What R-squared 0.031 actually means

For flagship models, the paper reports fitting an exponential decline curve and getting an R-squared of 0.031.

R-squared measures how much of the variation the trend line explains. A value of 1.0 means the trend explains everything. 0.031 means it explains about three percent, which is another way of saying there is no reliable downward trend to speak of. Flagship prices move, but not in a direction you can plan around.

This makes intuitive sense once stated. The frontier tier is defined by being the most capable thing available. When a model gets cheaper it stops being frontier and something else takes the position at a similar price. The tier holds its price because the tier is defined by position, not by a fixed capability.

So "AI is getting cheaper" and "the best AI stays expensive" are both true, and the second one is the one that hits your invoice if you built on the best.

The reasoning premium

The paper puts the reasoning premium at an average of 31.5x non-reasoning prices.

Reasoning models also produce far more tokens than they show you, because the intermediate reasoning is generated and billed. A 31.5x per-token premium on a model that generates several times as many tokens compounds into a very different bill from the one you estimated off the price page.

This is the most common budgeting error we see. Someone estimates cost from the published per-million-token rate, multiplies by expected volume, and gets a number that turns out to be low by an order of magnitude because they counted the visible output only.

Where the savings come from, and why it matters

The paper's most striking decomposition: software and architectural innovation account for approximately 103.7% of cost reduction, while GPU hardware contributes -0.9%.

The shares add to more than 100% because hardware moved slightly the wrong way, so software had to cover the whole decline and then a little extra on top. Essentially all of the price collapse came from better software: quantisation, distillation, serving optimisations, architectural changes.

That matters for forecasting because hardware improvement is relatively predictable and software improvement is not. If the declines were hardware-driven you could extrapolate with some confidence. Since they are software-driven, they continue only as long as researchers keep finding new efficiencies, and there is no schedule for that.

Anyone giving you a confident five-year AI cost projection is extrapolating a trend whose engine is unpredictable by nature.

What this changes about how you build

Start by separating the calls that genuinely need a frontier model from the ones that do not. Most business workflows have one or two hard judgment steps and a long tail of classification, extraction and formatting. Running all of it on the frontier tier is where budgets disappear, and it is the easiest thing to fix. The strongest version of this is to take the model out of those steps entirely, which we cover in stop prompting for the parts of your workflow that never change.

Then pick your tier partly on price trajectory, not only on capability. If an economy model does a task acceptably, your unit cost for it roughly halves every thirteen months while you do nothing. The same task on a flagship model costs about what it costs today in three years.

Finally, make the model swappable. The savings come from software innovation, not a predictable curve, so the cheap capable model of 2028 is probably one that does not exist yet. A system that routes every call through one interface can adopt it in an afternoon. A system with a vendor's SDK threaded through forty files cannot.

What we tell clients

We quote AI running costs as a range with the model tier stated, and we say plainly which direction we expect the range to move. On economy-tier work we tell clients the cost is likely to fall. On frontier-tier work we tell them to budget flat and be pleasantly surprised.

We also design for the swap. It is a small amount of extra work at the start and it is the difference between benefiting from the next efficiency jump and being locked out of it.

If you are budgeting an AI build, the number you want is not the published rate. It is the metered cost of a week of your real traffic, on the tier you intend to ship on. That is where our AI and automation estimates start, because it is the number that decides whether the project is worth doing at all. Ask us for it.