The leaderboard carried none of September's thirty model releases

The leaderboard carried none of September's thirty model releases

We keep a launch board, and every row on it records where the launch was verified from. Not where we first heard about it — where the dated evidence came from: the vendor's own announcement, its documentation, its pricing table, a model host, a forum thread, a press report, or the daily leaderboard that most people mean when they say they follow product launches.

That field was added because we kept getting dates wrong. It has turned into an unexpectedly useful record of where new software actually surfaces, so this month we counted it.

Read on 20 September 2026, our launch records hold 315 rows — 297 of them live on the public board, the rest our own announcements scheduled for dates that have not arrived yet. 118 rows are dated September 2026: 30 AI models, 87 products and one major update to an existing product. Every single one of those 118 carries a source.

72 of the 118 came from one launch leaderboard. Six of the 118 are our own products — three of them dated later this month — and we exclude all six from every figure in this piece; without them the count is 72 of 112, or 64%. Including our own rows it is 61%. Either way, roughly two thirds of everything we listed in September came from a single place.

And not one of the thirty model releases did.

What we counted, and what the categories mean

Each launch on the board carries a link to the artefact that established it: the page we read, dated, before the row was written. We sorted the September rows by the host of that link into five groups.

A launch leaderboard is a daily ranking site where makers post their own products. A vendor's own site is the company's announcement, its blog, its documentation or its pricing page — the place where a company states what it has released, on its own domain. A model host is a repository where weights and model cards live. Press is a dated report by a publication. A forum is a discussion site where a release was posted and argued about in public.

For the 118 September rows, the split is: 72 from a launch leaderboard, 33 from a vendor's own site, five from a forum, four from a model host and four from a press report. Excluding our own six, which are all our own product announcements: 72 from a leaderboard, 28 from vendors' own sites, five from a forum, four from a model host and four from press.

The reason to separate these is not tidiness. The five groups have different failure modes. A leaderboard tells you a maker chose to post today and tells you nothing about when the thing shipped. A vendor's own page tells you what the company is willing to put its name to, and usually carries a date. A model host tells you what was actually uploaded and under which licence, which is the only one of the five that is hard to overstate. Press tells you what someone else understood. A forum tells you what people who tried it think, which is worth more than any of the rest and is never a substitute for a date.

The models are announced somewhere else entirely

Thirty AI model releases were dated September 2026 on our board. Twenty-three were verified from the vendor's own site, four from a dated press report and three from a model host. None came from a launch leaderboard, and we read those leaderboards every day.

When you look at where the twenty-three vendor announcements actually live, the pattern gets sharper. They are not marketing pages. They are developer documentation and pricing tables: an API pricing page, a model guide, a developers' changelog. The release of a model is now, in many cases, a row appearing in a price table.

Two examples from this week make the point. The variant of a widely used fast model released on 18 September has, as far as we could find, no announcement post at all — no blog entry, no press release. What exists is the vendor's model guide, which lists it alongside its sibling with the same context window and the same maximum output, and the vendor's pricing table, which charges $0.37 per million input tokens and $1.25 per million output against $0.15 and $0.50 for the model it is built from. The release is a 2.5x price step for latency, and the only place it is documented is the page a buyer reads after they have already decided to use it.

The other is a 600-billion-parameter flagship announced on 20 September, whose API opened the same day. Its own documentation gives the context window and the modalities; its own pricing page gives $1.00 per million input tokens on a cache miss, $0.05 on a cache hit and $2.70 per million output tokens, with an explicit note that output tokens include the reasoning the model performs before it answers. That last sentence is worth more to a finance team than the entire announcement, and it appears only in the pricing documentation.

If your window onto new software is a leaderboard of things people posted today, the entire layer that determines what your software costs to run is invisible to you. You will see the applications built on these models within days. You will see the models themselves when your bill changes.

The date on the post is not the date of the release

The second thing the source field records is the size of the gap between the day a company shipped something and the day it surfaced where most people would see it. In this week's batch alone we dated three launches to the company's own announcement rather than to the post that brought them to our attention.

A coding agent that runs only on open-source models was announced by its vendor on 14 September, as a research preview running to 14 October, and reached the top of a launch leaderboard on 19 September — five days later. A full-text search extension for Postgres was announced by its vendor on 16 September, with a changelog entry of the same date, and reached a technical forum's front page on the 19th. A semantic-layer release inside an existing analytics product was posted to a leaderboard on 19 September and announced on the company's own blog the same morning, which is the only one of the three where the two dates agree.

A fourth is the reverse case. A legal-research configuration of a frontier model was announced by its vendor on 17 September and never appeared on any leaderboard we read. It is on our board because we read the vendor's announcement.

The gap matters more than it looks. Three of the products we listed this week carry dates that expire. The research preview above ends on 14 October, and the terms afterwards are asymmetric: subscribers on the more expensive plan lose access to the feature at the end of the preview, while subscribers on a cheaper standalone plan keep it. A buyer who believes the product launched on the 19th is five days wrong about a thirty-day window, and may be wrong about which plan to be on when it closes.

Dates also collide in the other direction. The 600B flagship is dated 20 September on our board because that is when its vendor announced it and opened the API. An independent benchmarking service lists the same model from 18 September, which is when that service measured it. Both dates are real and they mean different things. A model that a third party has already scored two days before its vendor announces it is a fact about access, not about release — and quoting the earlier date as the release date would be wrong.

The same week produced a third variant: a small open-weights model where two secondary write-ups gave 17 September and the vendor's own blog post, the vendor's own launch thread and the index that carries it all gave the 18th. We used the 18th. Three dated artefacts, two of them the company's own, against two undated write-ups.

What a single-source reader missed in September

Taking the September rows that did not come from a leaderboard, and leaving out our own, the list of what a leaderboard-only reader would not have seen is not a list of obscurities.

It includes every model release of the month: two frontier releases in the first three days, two image models on the 8th, a reasoning model and an engineering model on the 10th, an orchestration pair on the 11th, two live speech models on the 15th, a tabular foundation model on the 16th, an open-weights compression release on the 17th, an omnimodal release and a device-sized automation model on the 18th, and the 600B flagship on the 20th.

It also includes a specific class of product: infrastructure. The search extension for Postgres, the agent-facing search service priced by the second, the tool-calling command layer, the agent-experience checker, a sub-10-millisecond document engine built for European electronic-invoicing standards. Five products in ten days, none of which was on a leaderboard when we found it, and all of which a technical buyer would want to know about before their competitor does.

There is a reason for the split, and it is not that infrastructure vendors are shy. A leaderboard rewards a product that can be understood from one screenshot and tried in a minute without a contract. That description fits a note-taker, an outreach agent and a screen recorder. It does not fit a search index that has to be installed inside a database, and it does not fit a model that is consumed through an API key by people who read pricing tables for a living.

So the leaderboard is not failing. It is doing exactly what it is designed to do, for exactly the category it is designed for. The mistake is treating it as a complete picture of what shipped.

The blanks in our own record

Our own record is not clean, and the same count that produced the figures above also produced the evidence against us.

Of the 315 rows, 145 carry no source link at all. Those are older rows, written before we recorded where a launch came from, and their dates rest on whatever the person entering them had in front of them at the time. We cannot audit them the way we just audited September, and we are not going to pretend otherwise. The fix is not retrospective: from here, a row without a dated source does not get written.

The September rows carry a second admission. Of the 118, one was written on the same day as the date it carries and 109 were written between one and six days after it. Three were written a week or more later. We are not fast, and this piece is not an argument that we are. Our claim is narrower: the date on the row is the vendor's date, not ours, and where sources disagree we say which artefact we believed and why.

And six of the 118 were written before the date they carry. All six are ours — our own product launches are scheduled in advance, so their rows exist before the day arrives. They are excluded from every figure in this piece, and the one headline number changes from 61% to 64% when they come out, which is the only direction that disclosure could move it.

How to date a release without waiting for someone to post it

The practical version of all this is short.

Start at the vendor's own domain, and prefer its documentation to its blog. The blog says what the company wants understood. The docs say what the product does, and the pricing page says what it costs. This month, several releases existed only in the second and third of those.

Treat a leaderboard position as a date of posting, not a date of release. If the product matters to you, find the company's own announcement and use its date. Five days is a common gap and it is enough to matter when a preview window, introductory pricing or a plan change is attached to it.

Check whether a benchmark figure was produced by the vendor or by someone else. Both are worth having and they are not the same claim. Of the models on our board, only a handful carry a figure produced by an independent evaluator; the rest publish their own. A vendor's number measured against its own predecessor on its own harness is useful — it is just not a comparison with anybody else.

Look at the repository, not the announcement, for anything called open. A release this week announced open weights for a date three weeks out; the repository named in the announcement currently holds a single configuration file with no weights, no licence and no model card. The announcement is a statement of intent and the repository is the fact. They should not be read as the same thing.

Read what the terms say about your data before the usage allowance tempts you. One of this week's launches offers up to fifty times more usage on open models in exchange for opting in to have prompts, code, project files, configuration, tool calls, edit histories and fix traces de-identified and used to train open-weight models, with sessions from some jurisdictions excluded. That is a clear and honestly published trade. It is also a trade, and the reason we could describe it in that much detail is that the company wrote it down. Not all of them do.

None of this requires a subscription to anything. It requires knowing that the place where software is announced has separated from the place where software is discussed, and that for an entire category — the models and the infrastructure underneath everything else — the announcement now happens in a price table.