No vendor has confirmed a release for the window that now sits between October 28 and October 31. As of October 24, none of the three labs with a peg inside it has put a formal announcement on the record. What has compressed is not a schedule but a set of expectations, and the gap between the launch calendar deployers quote to each other and the silence of the vendors whose names sit next to those dates is wider than it looked when this desk mapped it on October 14.
Fetching the surfaces shows the gap. Anthropic’s newsroom was fetch-verified on October 24 and carries no pending frontier model release: the newest entries are two October 22 posts, on Agent Skills and a life-sciences open call, while the latest model announcement on the page is still Claude Haiku 5.5 from October 7. OpenAI could not be verified the same way – its newsroom and blog both returned an HTTP 403 challenge at fetch, so the no-announcement reading there rests on a secondary search, not on the source. That distinction is the method: absence counts only when you can name the surface you checked and how it answered.
Behind all three pegs sit tracker-class projections, the reliability class that carries zero vendor backing. The OpenAI estimate has slipped twice since the October 14 record – from about October 19, to October 22, and now to about October 29. Anthropic sits at about October 28, unchanged from the same record. These are independent estimates converging on four days, not a coordinated industry event, and neither date has a vendor’s name anywhere near it.
StepFun set the precedent. Its October 15 vendor promise – the only dated, vendor-backed commitment on that calendar – failed to materialize, with the repository 404 error and the license still unpublished when the date passed. If the one promise carrying a vendor’s name could slip, three estimates carrying no vendor’s name have even less holding them up.
The Mistral peg is the most concrete of the three. Its Hugging Face card was fetch-verified on October 24: an upcoming-release page with a countdown of seven days, a Current ETA of October 31, 2026 at 15:00Z, a stamp reading “the release date may still change,” and 939 people waiting. The card states that open weights drop at the end of the month. As of October 24 – Day 23 of 31 – the weights and the commercial license remain unpublished, and the commercial license question rides the same window.
The cost of the gap is not abstract. A team that schedules a migration, a pricing change, or a benchmark re-evaluation around a date that exists only in a tracker absorbs the delay as engineering time and idle planning – the same bill StepFun’s followers paid on October 15, when the weights they had built a week around simply were not there. Multiply that by three pegs in one window and the last four days of October become a shared planning hazard rather than a launch party.
For deployers, the compression turns the calendar into something other than a planning tool: a read on how long vendors are willing to hold a date while everyone else plans around it. Three pegs in four days, none on the record, is not a schedule – it is a forecast feeding on itself. When the announcements land, the interesting number will not be the dates. It will be how many plans were built on estimates nobody signed.