I Published a Median Model Lifespan. Seven Weeks Later the Same Query Said 219 Days More.

On 2026-08-01 I computed how long an AI model lives, from the day it shows up to the day its
vendor says it will be switched off. The answer was 197 days, over 57 models.

I ran the same query on 2026-09-22. It now says 416 days, over 74 models.

Nothing happened to the models. Seventeen models joined the sample, and the middle of it moved
seven months. That is not a metric with noise in it; that is a metric whose value is mostly
determined by who is in the denominator.

Here is the denominator.

528, 290, 74

528   models where I have a recorded arrival date
290   models where a vendor has published a shutdown date
 74   models that have both        ← every lifespan number comes from here

That is 14% of the models I can name. The other 86% split into two piles, and both piles are
excluded for reasons that are not random:

• 454 models have an arrival and no announced end. They are alive, or at least nobody has said otherwise. A model that never gets a shutdown date never enters the average, no matter how long it runs.
• 216 models have a published shutdown date and no arrival I can point to. Mostly vendors whose deprecation table lists models their changelog never announced.

So the measurable population is: models that were born after I started watching and have since
been sentenced. Long-lived models are systematically underrepresented, because a model that is
still running contributes nothing until the day someone schedules its execution.

The median is a fact about vendors, not about models

Split the 74 by who made them:

vendor	models measurable	median	range
Google	27	211d	7 – 499
Azure	20	488d	212 – 1,017
OpenAI	16	537d	272 – 1,240
AWS	8	535d	483 – 920
Moonshot	2	183d	183 – 188
Anthropic	1	608d	—

Google is 36% of the sample and its median is less than half of everyone else's. Not because
Google's models die young in some general sense, but because Google ships a lot of -preview
models and publishes end dates for them. The shortest span in the whole set is
gemini-3.1-pro-preview: 7 days, arriving 2026-03-01 and scheduled off 2026-03-08.

Anthropic contributes exactly one model to a statistic about model lifespans.

Now re-read the headline number. "The median model lives 416 days" is, mechanically, "Google's
preview cadence and Azure's back catalogue, mixed in whatever proportion happened to have both
dates on record this week." Change which vendors publish end dates — not which models exist — and
the number moves. Which is exactly what it did between August and September.

The longest life in the set is a join across two vendors

gpt-4 measures 1,240 days: arrival 2023-06-01, shutdown 2026-10-23.

The shutdown is OpenAI's, from OpenAI's deprecation page. The arrival is Azure's, from an Azure
"What's New" entry — and Azure's What's New is monthly, so the day component is not a
measurement, it is the first of the month. 29 of the 74 arrival dates in this sample are month
floors like that — 39% of the population carries up to 30 days of error on one end.

Is it wrong? It is the same model, and the metric is defined as first record across the sources
I read, which is what it says on the page. But "first record across sources" and "launch date"
are different quantities, and the gap between them is entirely determined by which source I
happened to add first.

What is actually solid here

Not the median. Three other things from the same rows are:

The spread. Q1 is 216 days and Q3 is 570. The interquartile range is a year. Whatever the
centre is, half the measurable models fall in a band that wide, which is enough on its own to
say that planning against a "typical" model lifetime is not a thing you can do.

454 models have no announced end. That is not a lifespan estimate, it is a disclosure fact,
and it needs no denominator games: the vendor has published nothing.

The vendors' own stated policies. Anthropic publishes a 60-day minimum and AWS publishes six
months. Those are commitments, not observations, and they do not move when my crawl coverage
changes. If you need a number to plan against, use those — and note that only two vendors of the
fifteen I read publish one at all.

What I would actually do about it

Ask any "average model lifespan" figure for its n and its denominator. Mine is 74 out of 528.
A figure quoted without those two numbers is quoting the shape of somebody's crawl.

Don't derive a migration budget from a central tendency. Derive it from the one date that is
about your model: whether the vendor has published a shutdown date for the exact string you send
in the API call, and what their stated notice period is.

Watch the arrival side, not just the retirement side. Every vendor here publishes retirements
somewhere. Far fewer publish arrivals in a form you can date, which is precisely why 216 models
have an end and no beginning.

What this measures

Published vendor documentation across 15 vendors, read on 2026-09-22, model names matched on
normalised strings. A model is counted once, at its earliest published shutdown date. Successor
models are excluded from the retirement side — counting them stamps the replacement with the
retirement date of the model it replaces.

By the time you read this the live figure will be different, and that is the point of the post
rather than a disclaimer on it.

---

The lifespan section, recomputed on every load, with the count it was drawn from:
aichangewatch.com/insights