The busiest category on our launch board is the layer between your app and the model

The busiest category on our launch board is the layer between your app and the model

Models get the headlines. On our own board they are also, now, the single largest category — and they are still outnumbered by the tools built to manage them.

In the 30 days to 12 September 2026 the TrustList launch board holds 71 launches. One is our own company's product, which we come back to below because it matters. Of the other 70, nineteen are AI models — a new model every day and a half, and comfortably the biggest single category.

But three categories underneath it add up to 21, and all three describe the same part of the stack: the plumbing between an application and whatever model it calls.

API management, nine launches. Application development, eight. IT management, four. Those are not model companies. They are the toolchain that decides which model gets called, what it is allowed to do, what it costs, and whether anyone noticed when it broke.

Twenty-one against nineteen is not a landslide. It is roughly level, and that is the honest way to put it. What makes it worth writing about is that the nineteen come from a handful of very large labs, and the twenty-one come from twenty-one different companies.

A disclosure, and a correction

Until earlier today this window contained nineteen launches from Rutba — our own company — all posted on 8 September, just under a quarter of everything here.

They were originally scheduled across two months, one or two a week. A decision on 8 September moved them into the board's Featured group and made them all live at once, and the direct consequence was nineteen entries from one vendor on a single day, burying eighteen of them and crowding every other maker out of that day's view. That has now been reversed: the suite is back on its original calendar, running from 10 September to 12 November, so only Rutba Office appears in this window and the rest return on their own dates.

We are leaving this paragraph in rather than quietly recomputing, because the first version of this article drew a conclusion from the clustered data that did not survive the correction. It read the long tail of single-launch categories as evidence of a market too fragmented to consolidate. Most of that tail was us.

What did survive, both corrections, is the top of the distribution. Our launches never touched AI models, API management, application development or IT management, so the four categories this article is actually about have not moved by a single entry through any of it.

A new model every day and a half

Nineteen models in 30 days is the number that reframes the rest.

It is worth being precise about what that figure is. This board records a model only when the release can be verified against a named, fetchable source — usually the vendor's own announcement, pricing page or model reference. Releases that exist only as a claim in an aggregator do not get recorded. So nineteen is a floor, not a census.

It is also accelerating. Five of those nineteen shipped in the four days to 11 September alone, from three companies. At that cadence a buyer evaluating models one at a time is running a race they cannot finish: by the time a procurement process has worked through a shortlist, the shortlist has changed.

This is the practical reason the interesting engineering has moved down a layer. If the model tier turns over every fortnight, the durable investment is in whatever lets you swap models without rewriting the application.

Which is precisely what the second and third categories on the list are.

API management is the biggest non-model category

Nine launches in a month, in a category that four years ago meant REST gateways, rate limiting and developer portals.

That is what "the layer between your app and the model" looks like when it lands in a taxonomy that predates the thing it is now used for. The work being launched here is routing, metering, key management, fallback, caching, cost attribution and policy — the concerns you acquire the moment you have more than one model, more than one team calling it, and a finance department that has noticed the bill.

Read alongside what the model vendors shipped in the same window, it fits uncomfortably well. Sakana AI released an orchestration system it explicitly describes as "an architecture, not a single model", which "dynamically routes tasks to the leanest model capable of solving them". DeepSeek released a model with two prices depending on the hour of day, and a cache rate fifty times below its cache-miss rate. OpenAI released a capable image model and a fast one and left the choice of which to call entirely to the caller.

Every one of those releases pushes work down into the calling layer. Somebody has to decide which tier to call, whether this workload can wait for off-peak pricing, whether the prompt is structured so the cache actually hits, and what happens when the routing decision is wrong. That somebody is increasingly a piece of API-management software, and nine of them launched last month.

The shape of the tail

With our own entry set aside, the picture is 70 launches across 22 categories, of which 11 contain exactly one launch.

Eleven singletons out of 22 is a scattered field, and it is close to what you would expect from any sample this size, which is why we are no longer building an argument on it. Content management, appointment scheduling, project management, help desk, team collaboration, risk management, recruiting, feature management — one launch each. That is either a market where nothing is converging, or seventy data points behaving like seventy data points. One month cannot tell the difference.

The middle of the distribution is more interesting than either end. Document management, business process management, compliance, business intelligence and sales enablement all sit at three. Compliance at three, in a 30-day window, is worth watching: it is the category where AI features create obligations rather than capabilities, and a steady trickle there is more consistent with real adoption than a spike anywhere else would be.

Almost nobody announces an update

One more number is worth pulling out, because it says something about what a launch board can and cannot tell you. Of the 70 third-party launches, 49 are new products, 19 are AI models, and exactly two are major updates to something already listed.

Two. In a month. Across a whole industry that ships continuously.

That is not because nothing was updated. It is because a launch board is understood, by the people who post to it, as a place to announce a birth. The second version of a product that has been in market for three years is a harder story to tell in that format, so it does not get told there, even when it represents far more engineering than a first release.

The consequence is a systematic blind spot, and it is worth naming because it applies to every board of this kind including ours. We can tell you what appeared last month. We cannot tell you, from this data, what survived. A category with nine launches and a category with nine launches of which seven are dead by Christmas look identical here. For a directory whose purpose is helping people choose software that will still be supported in two years, that is the more important question, and a launch feed structurally cannot answer it.

This is why the board is one input to a listing rather than the listing itself. What happens to a product after launch day shows up elsewhere — in whether it accumulates reviews, whether the company claims and maintains its listing, whether anyone is still asking questions about it six months later.

What "high value" means when nobody can trust the votes

There is a question hiding under all of this: what makes a launch worth paying attention to at all.

The usual answer on a launch board is votes. We deliberately do not use it. Every launch this board seeds from a verified source is recorded with a vote count of zero, because inventing popularity is the fastest way to make a directory worthless, and importing a vote count from another platform measures that platform's community rather than anything about the product.

That leaves us needing a different definition, and the honest one is duller than a leaderboard: a launch is worth attention if it can be checked. Does the vendor publish what it costs? Does it state a context window, a licence, a benchmark with a named harness? Is there a page you can open today that says what the thing is and what it charges?

By that standard the month's model releases score very unevenly. DeepSeek publishes a complete price table — peak and off-peak rates, cache rates, the exact hours the peak window covers — so you can calculate a bill before signing anything. Sakana AI publishes a price for one of its two releases and benchmark figures for the other, and not both for either. OpenAI publishes a one-line description of each image model and neither prices nor benchmarks.

None of that maps onto how good the models are, and it is not meant to. It maps onto how much of a purchase decision a buyer can make without booking a call. On a directory whose entire job is comparison, that is the more useful axis, and it is the one where the gap between vendors is widest.

How to read this, and where it breaks

This is one board's view of one month. The limits are real, and this article has already been wrong once.

The categories are the board's, not the market's. "API management software" is a taxonomy that predates model routing, and some of those nine launches would describe themselves differently. The category tells you where a thing was filed, which is a decent proxy for what it does and not the same as what it is.

Seventy third-party launches is a small sample. One busy week in a single vertical would move the ranking. Twenty-one against nineteen is a lead that a fortnight could erase.

The model count moves when we do our job. Five of those nineteen models were added to the board on the day this was written, during a catalogue refresh. That is the board getting more complete rather than the market speeding up, and it is a reminder that a count from a curated source measures the curation as well as the world.

A board records what it can verify. That biases toward vendors who publish properly and against products launched by people with no time for documentation. It is a bias we would choose again, and it is still a bias.

The board is young. It holds 221 launches all-time, so this single month is roughly a third of everything on it. Month-on-month comparisons will mean something in six months; today they would be noise.

What would change our mind

The claim here is that value is concentrating in the layer that routes, governs and meters calls to models. It is worth saying what would falsify it.

If API management and application development fall back to two or three launches a month while the model count holds, the concentration was a spike and we will say so. If the orchestration pattern in the model releases turns out to be one unusual fortnight — if the next dozen releases are single models with flat, time-independent pricing — then the argument connecting the two halves weakens considerably.

And if compliance and business process start climbing while API management flattens, the story changes shape entirely: it would suggest the binding constraint had moved from technical routing to governance and audit, which is a different market with different winners.

We will have the data either way next month, on the same endpoint, and we will report it whichever direction it points — including if it points away from what we have argued here.

Sources

• TrustList launch board, queried 12 September 2026 after the catalogue refresh and after our own suite was returned to its staggered calendar. Every launch, category and type count above is reproducible from this endpoint; the third-party figures are the same set with our one remaining entry excluded: https://backend.trustlist.uk/api/launches/board?window=month&limit=200&pageSize=200
• Sakana AI, "Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier", 11 September 2026: https://sakana.ai/fugu-max-release/
• DeepSeek, API pricing documentation — the peak/off-peak table, the peak-hours window and the cache rates: https://api-docs.deepseek.com/quick_start/pricing
• OpenAI, model reference — the GPT Image 2.5 Sunburst and Flare entries: https://developers.openai.com/api/docs/models