What 215 launches in ninety days say about where tooling is going

What 215 launches in ninety days say about where tooling is going

TrustList runs a launch board. Over the ninety days ending 8 September 2026 it recorded 215 launches: 129 products, 84 AI models, and 2 major updates to products already listed.

That ratio is the interesting number. Not because it measures the industry — a directory is not a census, and the caveats below are real — but because two different kinds of release are being tracked on the same board by the same standard, and one of them is behaving very differently from the other.

The model cadence is not a straight line

Counting AI model launches by the month they actually shipped, rather than the month anyone wrote about them:

• June 2026: 22
• July 2026: 35
• August 2026: 19
• September 2026: 8 so far, through the 8th

July was an outlier by a wide margin — thirty-five model releases in thirty-one days, more than one a day sustained across the month. August fell to roughly half that. Early September is running at a pace that would land somewhere between the two.

The obvious reading is that the release cadence is cooling. It is worth resisting that reading for a moment, because a board like this is sensitive to two things that have nothing to do with how many models exist. The first is what gets verified: every model on this board has to trace to a named, dated source, and releases that could not be verified to that standard were left off rather than estimated. The second is that "a model release" is not a fixed unit. A lab shipping one flagship and a lab shipping a flagship plus three size variants produce very different counts for arguably the same event.

What the numbers do support is narrower and more useful: the extraordinary compression of July was not the new normal. Four frontier labs shipping inside seventy-two hours, as happened at the start of September, still reads as unusual against the surrounding months.

Ninety days of product launches, and what they have in common

The 129 product launches are the other half, and the six added this week are a fair sample of what is arriving. All six were verified individually against their own Product Hunt pages and the domains those pages link to before being listed.

• Tadata — an agent that lives in Slack, holds team-wide and person-specific memory, and only participates when tagged.
• DocsAlot Visual Editor — a documentation editor whose visual mode does not corrupt the Markdown underneath, with review-before-publish drafts.
• Kit by Speakeasy — a coding-agent runtime as one static binary, speaking Agent Client Protocol and A2A.
• PR Lens by Coldtea.ai — animated architecture and data-flow diagrams for a codebase or a pull request, MIT-licensed.
• Routines by Databox — an analyst-agent that runs an analysis and delivers a report on a schedule.
• Airuncode — a local-first runtime for running several coding agents on your own machine, on your own API keys.

Three patterns run through them, and none is about model capability.

The unit being sold is the runtime, not the model. Kit and Airuncode both sell the thing that runs agents rather than the intelligence inside them. That is a market position that only makes sense once model choice is assumed to be fluid — you do not build a runtime abstraction over something you expect to pick once.

Ownership is a feature now. Airuncode's pitch is bring-your-own keys, switch between cloud and local models, pay providers directly rather than through a marked-up resale, and the maker states the goal as avoiding dependence on any single AI company, subscription or model provider. Tadata states it never trains on customer data and is model-agnostic. PR Lens is MIT-licensed. These are positions taken in response to a default, and the default is the opposite.

Scheduling and review are where the value moved. Routines is not a better analyst than a chat window; it is the same analysis on a cadence, delivered where people already are. DocsAlot separates editing from publishing so a human approves before anything goes live. Both are betting that the hard part is no longer generating an answer but putting it into an operational loop somebody trusts.

A fourth pattern is visible in what these products decline to do. Tadata only speaks when tagged. DocsAlot holds changes in a draft until a person approves them. PR Lens shows a reviewer what changed and leaves the judgement to them. Each of those is a deliberate restraint, and restraint is a strange thing to advertise unless the market has recently had too much of the opposite. Twelve months ago the competitive claim in this category was autonomy — how much the product would do without being asked. This week's launches compete on how precisely they can be told when to act, which is a different product and a different buyer.

That shift is consistent with what analysts are describing as the year's theme, and it is more persuasive coming from shipping products than from a forecast: an agent that interrupts a channel unprompted is a demo, and an agent that waits to be tagged is something a team will still be running in six months.

What the board cannot tell you

It is worth being explicit about the limits, because a number like "215 launches" invites more weight than it can carry.

This is a filtered sample, deliberately. Product launches on this board come from Product Hunt's organic daily leaderboards, filtered by a B2B bar that excludes consumer novelty, games, gambling-adjacent products and adult content. This week alone that removed a crypto trading agent, an iOS teleprompter and five desktop utilities. That filter is a judgement, and a different filter would produce a different 215.

Absence is not evidence. A product not on the board may simply not have launched on a channel the board watches, or may have failed a verification step for a reason that says nothing about the product — a name collision with an unrelated tool has caused exclusions before.

Vote counts here are not popularity. Every launch seeded by our own editorial process starts at zero votes, because fabricating engagement on a board that exists to be trusted would defeat the point. Rankings reflect what readers do after publication, not what we expected.

Ninety days is short enough to be seasonal. The window covers a European and North American summer, which is not a neutral quarter for enterprise software releases. July's thirty-five model launches and August's nineteen may say as much about release calendars and holiday schedules as about any underlying trend, and a single quarter cannot separate the two. The honest version of the cadence finding is that July was exceptional relative to the months around it — not that the industry is speeding up or slowing down, which would need a year of the same measurement taken the same way to support.

Why we publish the count anyway

A directory that only publishes conclusions is asking to be trusted. One that publishes the counts underneath them can be argued with, which is the more useful posture.

The specific thing worth taking from this quarter: the centre of gravity in tooling has moved from the model to everything around it. Eighty-four model releases in ninety days is a remarkable number in absolute terms, and it is still outnumbered three to two by launches of products that treat the model as a component they route around. The runtime, the schedule, the review step and the question of who holds the keys are where this quarter's products competed.

That is not a prediction. It is what 215 verified entries on one board looked like between June and September 2026.

The two models added this week, and why they were missing

The board gained two AI models on 8 September, both of which had shipped earlier and neither of which had been listed. How they were missed is a better illustration of how this board works than the count is.

Qwen3.8-Max-0902 is a dated snapshot of Alibaba Cloud's Qwen3.8-Max. The board already carried Qwen3.8-Max from 6 August; the update carries the vendor's own 09-02 in its name. Alibaba's Model Studio notice describes improvements in coding depth, agentic collaboration across multiple tools, and visual understanding, retains the 1M-token context window, and states that billing and pricing are unchanged.

It is also a good example of what does not get recorded. Third-party write-ups quote specific per-token prices for this snapshot and a table of benchmark wins against a competing frontier model. None of that appears in Alibaba's notice. So the listing carries no prices and no scores — a gap that is visible on the page rather than filled with a plausible number. There is a second discrepancy kept in the open: the notice gives an effective date of 5 September for Model Studio availability, while the snapshot is named for 2 September and was reported on that date. The launch is dated to the release, and the other date is stated rather than quietly discarded.

Claude Mythos 5.1 shipped on 1 September, the same day as Claude Fable 5.1, which the board already listed. Anthropic states the two are the same underlying model: Fable ships with safeguards for cybersecurity and biology and is generally available, while Mythos runs with those safeguards reduced for authorised users. Access runs through vetted programmes for cyberdefenders and life scientists, currently limited to a set of US organisations, and cannot be bought.

That last property is why it is worth listing separately. A directory that only carries the things you can purchase quietly implies that purchasability is the only axis that matters. A model with frontier capability and no commercial route to it is a real market fact, and a buyer comparing options benefits from seeing that the capability exists and is gated rather than not seeing it at all.

How an entry gets on this board

Because the numbers above are only worth as much as the process behind them, it is worth stating the process plainly.

Every item traces to a named, dated source. For models, that means the vendor's own announcement, pricing page or release notice wherever one exists, with an independent report as corroboration. For product launches it means the product's own Product Hunt page and the domain that page links to, opened and read, not a scraped title.

Name collisions are assumed until ruled out. This has caught real problems: in previous batches, products whose names matched an unrelated tool from years earlier were dropped rather than guessed at. A product that cannot be resolved to a specific page is left off.

Dates are the release date, not the discovery date. A model found three days late is dated to when it shipped. This is why the monthly counts above can still change slightly: a backfill lands in the month it belongs to, not the month it was added. Both models added this week were dated to early September, which is why September's count moved without anything new having shipped.

Figures that cannot be verified are omitted, not estimated. The visible cost of that policy this week was two prices and a benchmark table. The invisible benefit is that every number on the board can be followed to a source.

Nothing is pinned. Editorial ranking exists in the software and is not used for launches we seed ourselves. Everything starts at zero votes.

None of that makes the board complete. It makes it checkable, which is a different and more achievable goal — and it is the reason the caveats in the previous section are printed alongside the number rather than left for a reader to discover.

Sources

• TrustList's own launch board, queried 8 September 2026 — launch counts by type and by month, all dated to each item's actual release date.
• Product Hunt daily leaderboard, 6 September 2026 and 7 September 2026 — the source of this week's product launches, each then verified on its own product page.