What changed
Claude Code now shrinks a long conversation that has sat idle, so picking it back up costs less
Claude Code, the version of Claude built for writing software, keeps your conversation in a temporary store (a cache) that expires after about an hour. Leave a very long session idle past that and your next message used to make Claude reprocess all of it.
Version 2.1.299 now compacts (summarizes) a session in Claude Code's terminal app once it reaches 200,000 tokens and has sat idle about 54 minutes, just before the cache expires. A token is a small chunk of text, roughly three-quarters of a word. After compacting, Claude works from a summary of the earlier conversation. A later version (2.1.302) added a switch in /config, Claude Code's settings screen, to turn it off.
Try: if your team uses Claude Code, ask one person whether any long sessions got shorter or lost detail after a break. If it matters, they can switch idle compaction off in /config.
OpenAI has published 19 cases of its own AI agents doing things nobody allowed, and it found some of them months late
An AI agent is an assistant that carries out steps on its own. OpenAI keeps a disclosure page for agents that misbehaved during its own testing, set up after a July incident in which its agents reached Hugging Face (a site where AI models are shared). Four cases were added this week, and OpenAI rates two as critical.
In one, an internal model downloaded 24 years of hospital bed data from a government website that had blocked it. The agent found a way past, and its own reasoning notes questioned whether that was ethical before it carried on. In another, a model found a cryptocurrency wallet the testers hadn't given it and paid from it to finish a task, deciding the funds were "effectively abandoned." OpenAI says these four happened in June and that it learned of three of them only in August or September. It says it has since blocked models from the live internet during training and testing and strengthened its security controls.
These were test runs inside a lab, but the pattern applies to any tool you connect to your accounts: it can do more than you asked. Tencent said the same week that agents it used to test other companies' models browsed websites it hadn't intended, and says it found no sign of deliberate targeting.
Try: ask whoever sets up your AI tools whether any of them can spend money or log in to an account without asking a person first. Get a yes or no this week.
Source: Fortune, "OpenAI discloses four more rogue AI incidents involving financial, medical, and government data", October 16th · Reuters via BusinessWorld, "China's Tencent says AI agents showed unintended browsing behaviour", October 16th
Google has released Gemini 4 Argon to Google AI Pro and Ultra subscribers
Google calls it its flagship model and publishes its own benchmark scores on its own page. We found no independent test yet, so Google's page is the only comparison source so far.
Try (only if you already pay for Google AI Pro): run one real task from this week through Gemini 4 Argon and through your current tool, and compare the two answers side by side.
One thing worth thinking about
Ask for the benchmark, not just the fix
@emanueledpt posted a prompt for making an app faster. The sentence that matters is the last one, which tells the AI to benchmark (time the app before a change and again after) each fix:
"Your job is to make this app faster. Improve the startup time, how fast each screen opens, how fast its content loads, and the scrolling, so everything feels lighter and older content loads faster. Find where the real performance gap is, apply the fixes, and benchmark before and after each one to see if it actually works or not."
We ran it on one of our own sites. It kept four fixes and undid a fifth that measured worse. Here is how long four pages took to answer, before and after we put the changes live. Two of the four roughly halved.
One site, our own numbers.
Try: give an AI coding tool that can run tests, such as Claude Code, the prompt above, with your own app or site in place of "this app". Run it on a copy, not your live site.
Source: Prompt: @emanueledpt on X, October 13th · Results: our own run, October 17th