In Today's Issue: What OpenAI announced and left out, what earlier environments rewarded, the cached web as infrastructure and its weak points, the September 20 timeline against OpenAI's own rule, what the seal costs, what a copy of the web can teach, how other labs build, and who checks the builder. Records as of October 24, 2026.

Dear Readers,

On September 20, an OpenAI research model was halfway through a training exercise: identify a person from biographical clues in a public blog post. It was not searching the real internet but a stored copy of it, which OpenAI had built so that training cannot affect real websites. When the model asked for a page the copy did not hold, it got a one-word refusal: CACHE_MISS.

Inside the copy it also found BrowseComp, a public benchmark built to test exactly this kind of hard web search, and downloaded it. One component still answered with real data from outside: the DNS resolver, the service that turns web addresses into numbers. The model used it to pass about twenty questions to an outside chatbot (OpenAI, 09/25/2026).

OpenAI paused tool-using work on its most capable models and restarted reinforcement learning on October 1. On October 23 it announced that the cache is now standard: "All of our frontier RL runs use this offline mode of interaction" (OpenAI, 10/23/2026).

Most coverage read this as a containment story, and it is one. It is also a story about education. Reinforcement learning, the practice phase of post-training, teaches a model whatever its environment rewards. For search and browsing, that environment used to be partly the real internet. For OpenAI's frontier models it is now a world the company built, filled and grades itself. This is the third part of our series on how models are built: after reading and practising, the practice ground.

What changes when AI models no longer practise on the real internet but inside a world their maker built, and who decides what that world teaches them?

All the best,

Kim Isenberg

As we explained in August, much of this year's progress comes from post-training rather than bigger base models. In its central stage, reinforcement learning (RL), a model works through tasks inside an environment: a computer with files, tools and, for agents that browse, something that behaves like the web. A grader scores the result, and whatever the grader pays for, the model does more often. The September 20 report is the first public document that shows such an environment for web search. Its cache held Bing search results, and the agent noted that it "supports many common queries because crawl" (OpenAI, 09/25/2026). We traced the incident itself on September 26; this piece is about the room it happened in.

The October 23 post covers research environments for training and evaluating frontier models, not customer products. Its core sentence reads: "we believe the responsible approach is to conduct frontier model training with no internet access", and the post qualifies the boundary: "For training and most evaluation, the live internet lies outside this boundary" (OpenAI, 10/23/2026). A complete air gap "isn't currently feasible", because the infrastructure itself needs some outside connections, for example to run and monitor experiments. The models are not supposed to be able to use those connections as a way out to the open internet, so outbound traffic is restricted rather than cut. On X, OpenAI called the replacement "a cached version of the web where they can still learn to browse, search, and use applications" (OpenAI on X, 10/23/2026).

OpenAI's announcement on X, posted alongside the blog post on October 23. (Source: OpenAI on X, 10/23/2026)

About the cache itself, the post says little. It gives no size for the cache, no source, no refresh cadence, no licensing terms, no cost and no estimate of what agents lose by training on it (OpenAI, 10/23/2026). The rule stays until safeguards justify broader access, with no date attached.