The age discount: last season's model, this season's price
Model age is the quietest lever on an API bill. Across 207 dated models, the median price goes from $2.38/1M tokens under six months old to $0.8/1M past two years — 66% less for capability that was state of the art when it shipped. Every figure is today's cheapest live offer, and every dot below is one model, colored by the lab that made it.
Median price by age
Median blended $/1M tokens (cheapest live offer per model, $0 free tiers excluded), bucketed by time since release.
Every model, by age
One dot per model: its age today against its current price, log scale, one color per lab. Click a brand to isolate it; click a dot for its page.
The models don't get cheaper — the lineup changes
The gap between the bands above is easy to misread as a single model sliding down the scale over time. It is not what happens. 631 of the 866 endpoints under price tracking have not moved once in 144 days; where rates were revised, the count splits 1053 cuts to 834 rises, and the median rise adds 42.9% where the median cut takes back 20%. A dot barely moves after it lands. What changes is which dots exist — and which ones are still worth listing, which is the survivorship caveat below with a number attached. The dated log of the moves themselves is the changelog.
Ages computed from sourced release dates, prices verified 2026-10-23. Survivorship caveat: old models still listed are the ones providers still find worth serving — retired models don't appear. See also how fast models get replaced · the same model, priced differently · methodology. Age is one of 5 ways this market can be said to get cheaper, and they do not agree — the others are here.
FAQ
How much cheaper are older LLM models?
Across 207 dated models, the median blended price is $2.38/1M tokens under six months old against $0.8/1M past two years — 66% less. It is not the only spread this index measures: the same model costs different amounts from one host to the next, and the frontier has no natural price at all.
Does a model get cheaper as it ages, or is the market just moving?
The market moves; the model mostly does not. 631 of the 866 endpoints under price tracking have not moved once in 144 days, and where rates were revised the count splits 1053 cuts to 834 rises. So the gap between age bands above is composition, not accrual: it is which models exist and which are still listed that changes, far more than what any one of them charges.
Is an older, cheaper model good enough?
That depends on the task, and this index does not measure it — it prices models and benchmarks them, it does not evaluate them on your workload. What it can say is that the cheapest option and the most capable option are rarely the same model, and that the gap between them is what the efficient frontier is for.