Claude Code October Update: The Comments Are Asking for Models; the Team Fixed the Cache-Expiry Edge Case First

The Claude Code October update roundup is out. The official team listed a series of changes. The core is not a new model, but the friction in everyday use. First, here's the official recap video.

The official team says it updated mods, running multiple sessions side by side in the desktop app is smoother, and it provides more information about where usage goes. The mods point wasn't expanded on; for now, remember that desktop multi-session is the part prioritized for trial this time. A user named jd said desktop side by side sessions is the first thing he wants to try.

Idle Sessions Are Compacted Before Cache Expiry

The original explanation is direct: Claude Code now compacts long-idle sessions about an hour before the prompt cache expires. When you come back later, it only needs to reprocess the summary, not rerun the entire conversation.

For general readers, prompt cache can be understood as a temporary cache the model makes of the context. When the cache expires, resuming an old session often requires rereading a lot of content. Now, by compacting long-idle sessions into summaries first, you save usage and waiting time when you return.

When Pro and Max users resume extremely long conversations, if continuing would consume a large chunk of the 5-hour limit, Claude Code will warn first. This warning is more useful than checking usage after the fact.

Effort Can Be Adjusted for a Single Turn and Can Also Control Subagents

Typing ultrathink in a message now only raises the effort for that turn to xhigh; it won't change the settings for the entire session. If you want to think more and deeper about one hard problem, you don't have to drag all subsequent conversation along.

You can also tell Claude what effort level a given subagent should use. For example, if the main session is on high and you want the subagent for a certain task to use medium, just say so. A subagent is an instance Claude Code dispatches to execute a subtask, so setting effort separately is very practical for controlling costs.

In the CLI, /usage now explains in plain language what is consuming your limit. If there is a quick fix, press f to apply it.

High-Risk rm Still Prompts, Timed-Out Commands Move to Background

Even in bypass permissions mode, Claude Code will still ask first for a small number of high-risk rm commands. This is intentional. If you're in a disposable container and are willing to let these commands run without asking, you can set:

CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1

A user named Spinzu said that still prompting for high-risk rm under bypass mode is the quietest best change here. That judgment isn't hard to understand: everyday automation needs speed, but there should still be a boundary around deleting things.

After a shell command times out, it now moves to the background and continues running instead of being cut off directly. Long builds or test suites won't be killed halfway through.

Cloud Sessions, Diff Comments, and API Credit

Cloud sessions start setup while you're still typing the prompt, so there's less waiting after you send. In a cloud session's diff, you can click any line to leave a comment, and the comment is sent to Claude with your next message.

Max, Ultra, and Team plans now include monthly API credit for building things with the Claude API and Agent SDK. MCP servers that support this capability can ask you follow-up questions mid-task and can also run long tasks in the background. MCP can be roughly understood as a protocol that lets models connect to external tools and data.

The Comments Are Waiting for Models, the Official Team Is Polishing the Experience

Under the official recap, the main thread in the comments is very consistent: model releases. A user named Jadon said that seeing this, he thought Fable had been released and his heart almost stopped for half a beat. Archer said everyone only cares about when the next Opus or Fable will be released. Stats Wire said directly, you forgot to release Fable 5.5 in October.

Arpit Gupta posted a video with CLAUDE OPUS 5.5 Sep 22 written in the bottom right corner.

A user named embw_l0x also posted an image for Opus 5.5.

A user named Chris posted an exaggerated reaction video to that extremely long environment variable.

This update isn't explosive, but it will change the day-to-day feel. The official team is smoothing out the rough edges in usage, while the comments are waiting for the capability ceiling. Neither side is wrong; they're just on different rhythms. For the full update, see the official changelog: https://code.claude.com/docs/en/changelog.