Expected AI Model Releases

A lab's release gaps are heavily right-skewed — it ships a burst of models within days, then goes quiet for months. So the projection uses the median of the last five gaps rather than the mean: backtested across the full release history, the median predicts the next gap about 10% more accurately, and considerably better than that for bursty labs.

That median is added to the lab's most recent release date. Models shipped on the same day count as one release, so a lab that launches three sizes at once isn't credited with a three-day cadence.

Labs ship on working days, and not evenly across them. Fewer than 2% of dated releases landed on a Saturday or a Sunday, and Friday carries about half the releases an even split across the week would give it, while Thursday is the busiest day on the tracker. So every projection lands between Monday and Thursday, moving to the nearest of those days — and to Thursday when two are equally close. A date a lab has stated itself is left exactly as stated. The projection takes no account of which weekday a particular lab prefers: with only nine to twenty-eight release days each, no lab's favourite weekday can be told apart from chance.

Once a projected date passes without a release, the row shows how many days have gone by since it, struck through, and counts instead to a revised date: the soonest instance of the weekday that lab ships on most often. The row is listed under that revised date. Mistral has used more Wednesdays than any other day, so a delayed Mistral points at the coming Wednesday, and reads “Today” when that Wednesday is today. That is the one place a lab's own weekday is read, and it is not a sharper projection: the preference is far too weak for that, and a projection is not what it is for. It names the soonest day the lab plausibly ships, once the projected date has stopped being a date to expect. Candidates are Monday to Thursday as above, and a tie goes to Thursday.

Where a lab has publicly stated a date — its own post, an executive's own account, a livestream — that date is used instead of the projection. Leaks and third-party estimates are not used. A stated date that passes without a release stays on the list and shows as delayed, because a missed date is information too.

These are projections, not announcements. Treat them as a guide to which lab is due, not as a schedule.