When Notion released an MCP server (a connector that lets an AI assistant use an app’s data and actions), requests poured in fast enough to strain its systems and push the company to rethink its economics. That detail comes from a post Lenny Rachitsky published on X on 5 October 2026, listing his takeaways from a conversation with Tibo Sottiaux. It is the one concrete event in a list otherwise made of forecasts.
The layering matters here. Rachitsky is summarising Sottiaux, so every claim below is Sottiaux as paraphrased by Rachitsky, not Sottiaux’s own published words. The post discusses OpenAI’s hiring, its plugin programme, and a product called Dots in connection with the conversation. It never states Sottiaux’s title or employer, and this article does not supply one.
The Notion example carries the most weight. Rachitsky relays Sottiaux’s view that software agents will soon carry out most actions taken online, and that most people have not priced this in. Notion’s experience suggests the mechanism: a person clicks a few times a minute, while a program can call a service as often as its owner allows. A product built around human pacing can meet a very different kind of customer on launch day. The post gives no traffic figures, no date for the Notion launch, and no comment from Notion, so the size of the spike is unknown. “Soon” is not defined either.
The second claim is that the model picker, the menu where users choose a model and how hard it should reason, is likely to disappear. Sottiaux, per the summary, says he tires of making that choice himself and wants it gone, and that Dots ships without one. Removing the menu does not remove the decision. It moves to the vendor, which then chooses which model answers and what that costs. A user who can no longer see the choice also cannot see the trade-off between quality and price. That consequence is our reading; the post does not discuss it.
The third claim is the one most likely to be repeated. Rachitsky writes that builders should plan as if models will be roughly ten times better within a year, and that Sottiaux expected cheaper and faster models at rates the post calls “quite incredible”. The post never says what ten times better means. Ten times cheaper, ten times faster, ten times more accurate, and ten times longer context are four different bets, and a team could pick the wrong one. No benchmark, price list, or release schedule backs the figure.
The post also admits a miss on Sottiaux’s side: reality outran his own forecast, because he had pencilled in the current level of capability for a year or two later. That is a self-reported record with no details attached. It shows that he is confident in his predictions, which is different from showing that they are accurate.
Taken together, the three claims share a shape. Each is a forecast about the near future, offered in conversation, relayed through a second person’s bullet points. None is a measurement. Only the Notion episode describes something that has already occurred, and even that arrives without numbers.
For a product team, the useful step is cheap and does not depend on any forecast being right. Check what your rate limits, pricing, and infrastructure do if a large share of your callers turn out to be software acting for someone else. If the answer is that the system buckles the way Notion’s reportedly did, you have a launch-week problem regardless of whether the ten-times claim ever comes true.
Lenny Rachitsky on X, 5 October 2026, summarising his conversation with Tibo Sottiaux; the remarks are Rachitsky’s paraphrase, not Sottiaux’s published words.
