Half the class of 2023 did not reach its second birthday
“Is this project still alive?” is the wrong question in a field where most projects are new. Measured at the same age, cohort by cohort, two-year survival has fallen from 94.2 % to 55.5 % in seven years.
TL;DR — the short version
94.2 % of the projects born in 2016 were still being pushed two years later. For the class of 2023, it is 55.5 %. Nearly half of one year’s intake stopped before its second birthday.
The fall is recent and steep. Two-year survival sat between 89 % and 94 % every year from 2016 to 2020, dropped to 82.2 % in 2021, 67.8 % in 2022, and 55.5 % in 2023.
The first year got dangerous too. One-year survival was 97.8 % for the class of 2016 and 75.8 % for the class of 2024: one new project in four now stops within twelve months of being created.
What follows for you: a project’s age is now a real signal. Something started in 2023 that is still moving has already passed a filter that half its cohort failed.
“Is this project still alive?” is a question about today. It is also the wrong question to ask of a field that has quadrupled its intake in three years, because the answer is dominated by the fact that most projects are new, and new projects are alive by construction.
The question that actually tells you something is a cohort question: of the projects started in a given year, how many were still being pushed N months later? Everything is measured at the same age, so nothing is flattered by being young. Here is the catalogue by birth year.
Half the class of 2023 did not reach two years
1,816 projects in the catalogue were created in 2023 — the largest intake until 2026. 809 of them had stopped receiving pushes before their second birthday. Apply the same measurement to the class of 2016 and the figure is 18 out of 312.
This is not the usual “open source is abandoned” observation, which mostly measures the fact that old software stops changing when it is finished. It is the same test applied at the same age to every cohort, and it says that a project started in 2023 was roughly eight times more likely to be dead at twenty-four months than one started in 2016.
The first twelve months, which is where it starts
| Born in | Projects | Alive at 12 months | Alive at 24 months |
|---|---|---|---|
| 2016 | 312 | 97.8 % | 94.2 % |
| 2017 | 521 | 96.7 % | 92.7 % |
| 2018 | 562 | 97.2 % | 91.3 % |
| 2019 | 596 | 96.6 % | 91.4 % |
| 2020 | 614 | 95.4 % | 89.6 % |
| 2021 | 535 | 93.3 % | 82.2 % |
| 2022 | 734 | 87.6 % | 67.8 % |
| 2023 | 1,816 | 79.6 % | 55.5 % |
| 2024 | 1,231 | 75.8 % | — not yet 24 months old |
The twelve-month column can be read one year further, to the class of 2024, and it is the more alarming of the two. 298 of the 1,231 projects created in 2024 — 24.2 % — had stopped within a year of being created. For every cohort from 2016 to 2020 that figure was between 2.2 % and 4.6 %.
What changed
Three things happened at once, and the data cannot separate them, so we will not pretend otherwise.
The intake changed composition. 2023 onwards is when the catalogue filled with wrappers, demos, agent scaffolds and reproductions — projects built in a weekend around a model that was replaced within the year. A wrapper around an API that no longer exists does not get “abandoned”; it gets finished by obsolescence.
The cost of starting collapsed. When a working repository is an afternoon’s work rather than a month’s, the marginal project is far more speculative, and speculative projects stop. The same tooling that produced the agent wave also made it cheap to abandon.
And the field moved fast enough to invalidate work. A project pinned to a model generation, a prompt format or a plugin protocol that was superseded has nothing to push.
Method
All figures come from one snapshot of the olud.ai catalogue taken on 1 November 2026 at 00:03 CET: 11,394 projects carrying 63,822,390 stars between them. The catalogue is a curated subset of open-source AI on GitHub, not a census of it, so every share quoted is a share of that catalogue and not of the whole forge. Star counts move daily and the figures here will drift; the live numbers are on each project’s page in the catalogue.
A project counts as alive at N months if the interval between its creation date and its most recent push is at least N months. A cohort is only reported at a horizon once every project in it has had time to reach that age: the 24-month figures therefore stop at the class of 2023, and the 12-month figures at the class of 2024. This is a survival measure, not an activity measure — a project that ran for three years and stopped last week counts as alive at both horizons, correctly, because it survived them. Projects whose reported last push precedes their creation date (a handful of imports) are excluded. Cohorts with fewer than thirty projects are not shown.
Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.
Keep reading
- 42 % of the catalogue has gone quiet — the same question asked of the present instead of the cohort.
- 127 repos above 10,000 stars are held up by fewer than ten people — who has to keep showing up.
- 289 accounts hold half the stars — and 7,215 owners have published exactly one thing.