The year the product became a feature

Between the first of April and the end of July 2026, this archive filed eight new death notices. That is an ordinary stretch by volume. What is not ordinary is the cause of death. Almost none of these products ran out of money the way the 2023 cohort did, and almost none of them failed at the task they were built for. Huxe generated the audio it promised. Pixel Studio generated the images. GitHub Models served the models. Notion Mail sorted the mail. They worked, and they were switched off anyway, and in nearly every case what killed them was not a rival product but a capability that appeared inside something the user already had open.

Huxe is the cleanest specimen. Raiza Martin, Jason Spielman, and Stephen Hughes had worked on NotebookLM since its inception at Google, including the audio feature that turned a folder of documents into two synthetic hosts discussing it, and in December 2024 they left to sell that trick as a company of their own. Huxe read a user's mail and calendar and returned a daily briefing performed by AI hosts who could be interrupted mid-sentence. It raised 4.6 million dollars from Conviction, Genius Ventures, Dylan Field, and Jeff Dean, opened to everyone in September 2025, and lasted about eleven months. In that window Spotify, Adobe, Amazon, and Meta each shipped generated audio inside an application their users already had installed. None of those versions had to be better than Huxe. They only had to be closer.

Google ran the same play on itself. Pixel Studio arrived in August 2024 as a reason to buy a Pixel 9: type a prompt, get an image, keep a library, and later make stickers and edit generatively. It was a capable app with one structural problem, which is that the same company was building the same capability into Gemini, where it would reach every phone rather than one line of them. Google confirmed the ending to 9to5Google in February 2026 and carried it out on June 5 through the app's own version 2.3 update, which is a candid way to go: the last thing Pixel Studio ever did was install the notice saying it was finished. The app now shows a button pointing at Nano Banana in the Gemini app.

GitHub Models died the corporate version of the same death. It opened as a limited public beta on August 1, 2024 with an argument that was hard to dispute, namely that the largest population of developers on earth was already logged into GitHub, so the shortest path from curiosity to production ran through a free playground sitting next to the code. It shipped with Llama 3.1, GPT-4o, Phi 3, and Mistral Large 2, and a written promise that no prompt or output would be shared with model providers or used for training. By 2026 Microsoft wanted durable AI work running through Copilot or Foundry and did not need a third front door. GitHub closed the product to new customers on June 16, 2026, announced full retirement on July 1, and removed the playground, the catalog, the inference API, and the bring-your-own-key layer together on July 30. Every model in the catalog went on being available somewhere else.

Notion Mail is the strangest entry in the set, because the feature that ate it was built by the same company, in the same building, on purpose. Notion bought Skiff, spent a year turning it into an email client that sat on top of Gmail and did the sorting, and shipped it broadly in April 2025. It worked. Then Notion's agents got good enough at handling email end to end that more than half of Notion Mail's users were managing their mail without ever opening the inbox. That is a remarkable engagement number and a fatal one: an email client nobody opens is a rendering engine with a support budget. Notion announced in June 2026 that Mail closes on September 22, with the agents inheriting the workflows.

Relay.app shows the category-wide version. Jacob Bank sold his scheduling company Timeful to Google in 2015, ran Gmail, Calendar, and Workspace product for years, and left in July 2021 to build the sensible middle of workflow automation: the integration count of a Zapier, with a human approval step wired into the handoff so software about to do the wrong thing on your behalf had to ask first. Khosla Ventures led a 5 million dollar seed in October 2022, Andreessen Horowitz added 3.1 million a year later. Then every automation tool on the market became an agent company inside eighteen months, and the careful approval step went from a differentiator to a checkbox everyone shipped. Relay stopped taking signups on July 16, 2026 and posted a notice that opened with the words difficult news and never got to a reason.

Two more went to the acquirers whose products had been displacing them. Continue built an open-source coding assistant that plugged into the editor, ran against whichever model you preferred, and kept its configuration in a file you controlled, and it collected a Y Combinator batch, more than thirty-four thousand GitHub stars, and 5.1 million dollars without ever finding the part of that arrangement anyone would pay for; Cursor took the team in June 2026 and deleted the hosted service's data on July 15. Bud, which began as Orchids in Y Combinator's Winter 2025 batch, promised that anyone could describe an application and get a deployed one, while Figma, Lovable, Replit, and Base44 shipped the same convenience on the same schedule; Figma took the team on July 7, 2026 and gave users eleven days before everything hosted went dark. The engineers went to work alongside Figma Make, the product that had been eating theirs.

Cove had already run the same course a quarter earlier, and it is worth putting beside them because the founders were as well credentialed as this pattern gets. Stephen Chau, Andy Szybalski, and Mike Chu had built Google Maps features including Street View, and two of them helped launch Uber Eats. Their bet was against the chat transcript: Cove answered you on an infinite whiteboard, in cards and tables you could pick up and rearrange, and Sequoia led six million dollars behind it in 2024. What Cove never established was why the work had to happen on a canvas, and while it looked for that answer the general-purpose assistants absorbed the job. Microsoft hired the team in March 2026, the product closed on April 1, and every account was deleted.

The mechanism underneath all eight is not mysterious, and it is not really about artificial intelligence. It is that the marginal cost of adding a capability to a product with existing users fell below the cost of acquiring a user for a product that only has that capability. When a model is an API call away, a feature is a sprint, and the incumbent has the distribution, the billing relationship, and the habit. The startup has the head start and the better version. In a market where the head start is measured in months and the habit is measured in years, the better version loses. This is what commoditization looks like from the inside: not a price war, but a quiet realization that nobody has to open your app to get what your app does.

It is worth naming what makes this different from ordinary competitive failure, because the archive is full of ordinary competitive failure. When Neeva lost to Google, it lost to a better-funded product doing the same job in the same place. When Huxe lost to Spotify, it lost to a worse product doing a narrower version of the job in a place the listener was already standing. The winning feature usually is worse. It has fewer options, less polish, and none of the founder's taste. It wins on location, and location is the one advantage a two-year-old company cannot out-build. That is why the deaths in this cohort read less like defeats than like evictions.

The honest complication is that some of these businesses were thin to begin with. Huxe was, functionally, one feature with a company around it, and the founders had every reason to know that, having built the same feature inside Google. Bud was a convenience in a category with four well-capitalized versions of the same convenience. A fair reading says the commodity risk was visible from the first pitch deck and the market simply arrived on schedule. But GitHub Models was not thin, and neither was Notion Mail; both were competent products inside healthy companies, killed by strategy rather than by economics. The pattern is not only that weak products get absorbed. It is that in 2026 being absorbed became the default ending for a working product that lived one layer away from where the user already was.

What survives the pattern, on the evidence in this archive, is the thing that cannot be reproduced as a feature by the platform underneath it: a proprietary data position, a regulated relationship, a physical object, a workflow so specific that adding it to a general product would make the general product worse. Everything else is renting distribution from a landlord who is reading your product roadmap. Continue understood this and chose the open license anyway, which is why the code survives and the company does not. That is not a mistake so much as a trade, and the terms of the trade are now visible in the ledger.

The lesson the summer of 2026 leaves is narrower than the usual one about moats, and more uncomfortable. It is not that these products were bad, or early, or badly run. Several of them were built by the exact people who had invented the capability at a larger company and left to sell it properly. It is that the question worth asking before the first line of code is not whether you can build it, or even whether anyone wants it, but whether there is already a screen the wanting happens on. If there is, and it is not yours, you are not building a company. You are prototyping somebody's next release note.