Reading notes: how to read AI video changelogs
The notes cover why two logs give two dates, what a snapshot changes, how to read a platform's arrival and what a missing note does and does not mean. As of 2026-09-25.
Each note takes its examples from the register.
No model was run; the notes rest on the notes.
1In this section
- A platform's arrival date is not the maker's release date — Runway, Higgsfield and Luma announce outside models in their own changelogs.
- A snapshot is a release even when the name stays — When a maker moves a model name to a newer build, nothing in a user's request changes, but the output can.
- A user guide's document history is a changelog too — Amazon's Nova user guide ends with a document history table: each row a change to the guide, with a description and a date.
- Blog posts as release notes — Adobe and Amazon do not keep a changelog for their video models.
- Claims about quality and controls you can use — Some notes describe a new control: a parameter, a mode, an input.
- Features often arrive on one tier first — A new feature in a note often comes with a condition: pro mode only, paid plans only, Pro and Mega plans only.
- Four different numbers called clip length — Notes use length in four ways.
- How to read a model line's page — A line page lists every entry for one maker's models, oldest first, with four kinds of source: the maker's own note, a platform's arrival note, an undated…
- Terse logs and wordy posts — Makers write notes for different readers.
- What a missing note does and does not mean — No note is not the same as no change.
- When a model's feature lives in another product — Some changes reach users through a product other than the model's own.
- When names in the notes do not match — The same model can appear under several names.
- Why one change can carry two dates — Google publishes Veo changes in two places, the Gemini API changelog and the Vertex AI release notes, and they do not always agree on the day.