Skip to content
unerrunerr

AI got cheaper. Your bill didn’t.

In our own panel of 2,886 AI model SKUs over 35 months (Sep 2023 to Jul 2026), a model that is already live keeps its price in 99.2% of months. Almost nothing you can cut sits in the bill. What moves the number is which model runs which job, and you cannot make that call by reading an invoice.

Where the money actually is

Prices barely move. In our panel of 2,886 AI model SKUs over 35 months (Sep 2023 to Jul 2026), a live model kept its price in 99.2% of months. When one did move, it went up 42% of the time. The falling headline number is mostly new models arriving, not old ones getting cheaper.

So the saving is in switching. Moving a job onto a cheaper model is where the money is. That is a decision somebody has to make. It is not a discount that turns up on its own.

Switching blind costs more than it saves. PointFive (arXiv:2607.12161) cut tokens by 38% and total cost rose 6.8%. Patch success fell from 27 of 40 to 15 of 40. A cheaper model that needs three more attempts is not cheaper.

Which is why you have to see the work. A bill shows price and work delivered multiplied into one number, never the two apart. You cannot tell a cheaper model from a worse one by reading it.

What published prices did, 2023 to 2026

AI Price Index · Sep 2023 – Jul 2026

Index · 100 = Sep 2023 · log scale

Matched SKUs — the models you already runMeasured · NBER WP 34608−0.3%/yr−0.8% over 35 months
Quality-adjusted — the models you’d have to switch toMeasured · Gundlach et al.5–10× cheaper/yr96× to 681× over 35 months
AI price index by month, Sep 2023 – Jul 2026. Index 100 = Sep 2023.
MonthMatched-SKU index (100 = Sep 2023)Quality-adjusted index, 5× per year (100 = Sep 2023)Quality-adjusted index, 10× per year (100 = Sep 2023)
Sep 2023100100100
Oct 202310087.482.5
Nov 202399.976.568.1
Dec 202399.966.956.2
Jan 202499.958.546.4
Feb 202499.951.138.3
Mar 202499.844.731.6
Apr 202499.839.126.1
May 202499.834.221.5
Jun 202499.829.917.8
Jul 202499.726.214.7
Aug 202499.722.912.1
Sep 202499.72010
Oct 202499.717.58.25
Nov 202499.715.36.81
Dec 202499.613.45.62
Jan 202599.611.74.64
Feb 202599.610.23.83
Mar 202599.68.943.16
Apr 202599.57.822.61
May 202599.56.842.15
Jun 202599.55.981.78
Jul 202599.55.231.47
Aug 202599.44.571.21
Sep 202599.441
Oct 202599.43.50.825
Nov 202599.43.060.681
Dec 202599.32.670.562
Jan 202699.32.340.464
Feb 202699.32.050.383
Mar 202699.31.790.316
Apr 202699.21.560.261
May 202699.21.370.215
Jun 202699.21.20.178
Jul 202699.21.050.147

The line that barely moves is the models you already run. The line that falls is the models you would have to switch to.

Method and sources

Repricing is rare — a dated jump, not a trend: 0.82% of model-months carry a list-price change, and 42% of the changes that do occur are increases.

Matched SKUs — measured. Within-model price trend ≈ −0.3%/year, statistically indistinguishable from zero (Demirer, Fradkin, Tadelis & Peng, NBER WP 34608). Our panel of 2,886 model SKUs over 35 months shows the same rigidity from the other side: a live SKU keeps its price in 99.2% of months.

Quality-adjusted — measured, published, not ours. The cost of a fixed capability level falls 5–10× per year (Gundlach et al., arXiv:2511.23455); the band is that published range compounded across the window. A second estimate puts the median near 50× per year (Cottier, Snodin, Owen & Adamczewski, Epoch AI, 12 Mar 2025) — we chart the conservative one. The vertical axis is logarithmic: the two series span three orders of magnitude.

Two kinds of AI spend, two ways in

What your coding agents do, we measure ourselves. The instrument sits inside the agent’s loop and records the task, the turns, the model, the tokens, and whether the change survived the week.

Everything else is read from what you already have: document processing, classification, extraction, the assistants your teams run. We use your billing export at model granularity, your contract terms, and our panel of 2,886 AI model SKUs across every vendor.

One decision, two ways of seeing it. We would rather tell you which is which than blur them.

We don’t pick your models for you

Tools that swap a model and tell you quality held are guessing. Nobody can see that from a bill.

We measure what your current choices cost in work delivered — whether the change survived, and how often someone had to go back and fix it. Then you decide, on evidence.

What that gets you today

On the coding slice, unerr cuts 86-90% of the tokens an agent spends navigating code, measured against grep-and-read, with any “saving” that lost the answer thrown out. Reproducible on your own repo.

That number is a receipt, not the reason. The reason is the decision above it.

Run it for two weeks. The numbers are yours to keep.

Whatever it finds is your data, on your own machine. Uninstall and it stays.

curl -fsSL https://unerr.dev/install | sh

No account. Runs on your machine. Uninstall with one command.

On Windows: