Claude Code's October 2025 Update: 23 Releases, Explained
Claude Code shipped 23 updates in October 2025, from Haiku 5.5 to idle compaction to background agents. Here's what each one changes.
What changed in Claude Code during October 2025?
Anthropic pushed 23 versioned releases of Claude Code in October, moving the tool from version 2.1.293 to 2.1.399. Most of these updates are small enough that they never got individual announcements, but a handful change how the tool behaves on a daily basis. The headline additions are a new cheaper model called Haiku 5.5, a fix that makes the “Ultra Think” keyword actually raise reasoning effort again, a cost-saving feature called idle compaction, and a batch of improvements to how background tasks and sub-agents run. The desktop app’s Code tab also picked up several of these changes, sometimes later than the terminal version.
TL;DR
- Haiku 5.5 launched October 7th as the new default Haiku model in Claude Code, with a 1 million token context window and pricing of 10 cents per million input tokens and 50 cents per million output tokens (rising to 50 cents and $2.50 past 100,000 tokens in a prompt).
- Ultra Think now reliably increases effort on Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1, can be typed mid-turn to raise effort retroactively, and only triggers its visual “rainbow” indicator when it will actually do something.
- Idle compaction summarizes long conversations before the prompt cache expires, so resuming a conversation over 200,000 tokens after about 54 minutes of inactivity uses a short summary instead of reprocessing the full history.
- Sub-agents now run in the background by default in several contexts, and slow MCP tool calls no longer block the main turn while they execute.
- New permission controls support unattended runs, including a setting that denies rather than prompts for risky tool calls and a lock that blocks writes outside the working folder in every permission mode.
- The desktop app’s Code tab got dedicated fixes, including always-backgrounded sub-agents, a safer response to risky shell commands in bypass mode, and repairs to several longstanding bugs around MCP server disconnects and environment variables.
- The standalone Claude design site is shutting down December 14th, and Anthropic launched a separate proactive-agent app called Claude Duet on October 21st as a research preview.
What does Haiku 5.5 actually do differently?
Haiku 5.5 became the default Haiku model in Claude Code on the Anthropic API starting October 7th. It carries a 1 million token context window, and its pricing sits at 10 cents per million input tokens and 50 cents per million output tokens for prompts under 100,000 tokens. Once a prompt crosses that threshold, pricing shifts to 50 cents per million input tokens and $2.50 per million output tokens.
The practical reason this matters: Haiku isn’t typically the model answering your main questions. It’s the model running quiet background jobs inside Claude Code, like quick file searches or simple sub-agent tasks. Swapping in a faster, cheaper Haiku doesn’t require changing any settings. It just makes those background operations less expensive and, in many cases, faster, without touching the main model you’re chatting with.
How does the fixed Ultra Think keyword work now?
Typing “Ultra Think” in a prompt tells Claude Code to run that turn at extra high reasoning effort. After the October fix, this now functions correctly across Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1. One notable addition: you can type Ultra Think while Claude is already mid-task, and it raises the effort level for the remainder of that turn rather than requiring a fresh prompt.
Anthropic also cleaned up a cosmetic issue. The rainbow visual effect that signals Ultra Think is active now only appears when the keyword will genuinely raise the effort level, so it won’t falsely trigger if the phrase shows up inside a quote or a file path. For anyone who wants it off entirely, there’s an environment variable to disable the behavior.
A related fix addresses sub-agents. Previously, a sub-agent running on a different model could inherit your main chat’s effort setting inappropriately. Now a sub-agent uses its own model’s saved effort level by default, and a new environment variable lets you set what effort sub-agents run at independently. That means running your main conversation at max effort won’t silently push every sub-agent to max effort too, which helps avoid burning through usage limits faster than expected.
What is idle compaction and why does it save money?
Claude Code holds your conversation in a prompt cache for roughly an hour. If you step away from a long conversation and return after that window closes, your next message has to reprocess the entire conversation history from scratch, which consumes a meaningful chunk of your usage limit.
| Vendor | Tier | Risk | ApprovalNEW |
|---|---|---|---|
| Northwind Supply | Tier 1 | Low | Approved |
| Acme Logistics | Tier 2 | Medium | Approve |
| Helios Data | Tier 1 | High | Pending |
| Granite Freight | Tier 3 | Low | Approved |
| Pinecrest Labs | Tier 2 | Low | Approve |
Idle compaction addresses this directly. If a conversation exceeds 200,000 tokens and sits idle for about 54 minutes, Claude Code automatically summarizes it just before the cache would expire. When you come back and send a new message, it works from that compact summary instead of the full transcript. The feature started in the terminal and has since rolled out to VS Code and the desktop app’s Code tab. It can be turned off through the idle compaction setting in the /config menu, which matters if you need Claude to retain exact details from a long session rather than a summarized version.
Alongside idle compaction, Anthropic added a warning system for usage limits. If you’re on a Pro or Max plan and resume an old conversation that would eat a large portion of your 5 hour limit (previously called the “session limit,” now renamed for clarity), Claude Code warns you before proceeding and lets you choose to resume anyway or start fresh. The /usage command was also expanded to flag things that quietly consume tokens, such as max effort settings, background workflows, always-on context, and large tool results.
How has background task handling improved?
Several updates target the friction of running slow or long tasks. When an MCP server executes a tool call as a task, Claude Code now moves it to the background immediately and notifies the main conversation once it finishes, so a slow tool no longer freezes the whole turn.
The /tasks interface added an “earlier in this session” row, letting you reopen a completed sub-agent and send it a new message even after restarting Claude Code. Messages sent to background agents also arrive faster now, because a running command shifts to the background instead of forcing you to wait for it to finish first. And for long-running shell commands, background sub-agents automatically move a command to the background after three minutes, which prevents the prompt cache from expiring mid-build.
What changed around permissions and unattended runs?
A new /automodesetup command scans your project and recent sessions to configure auto mode with context about your specific setup, so it knows which actions genuinely need your approval. A “deny instead of prompting” setting lets a tool call that would normally wait on you get denied automatically instead, letting long unattended runs continue without a human present to click yes.
There’s also a setting that blocks all writes outside your designated working folders across every permission mode, adding a consistent safety boundary regardless of how permissive your other settings are. On the privacy side, Claude Code now asks for confirmation before a web or shell command sends your account email address over the network.
What happened with the desktop app and other Anthropic products?
The Code tab in the Claude desktop app received a cluster of fixes and features this month. Idle compaction arrived there as well. Sub-agents in the desktop app now always run in the background, matching terminal behavior, and “fork sub-agents” are on by default, allowing Claude to split off a full-context copy of itself for a task. In bypass permissions mode, a potentially dangerous rm command is now denied immediately with a suggestion to rewrite it, rather than halting an unattended session to ask for confirmation. The /code review command run from a cloud session in the Code tab now produces a structured findings card instead of a plain list.
Anthropic also fixed several longstanding bugs: project MCP servers disconnecting after the first turn, commands failing to read variables from a .zshrc file, and images attached mid-conversation being sent at full size, which had been causing errors in long chats.
Two other product notes from October: Anthropic launched Claude Duet on October 21st as a research preview for Pro and Max subscribers, a separate app where Claude operates as a proactive agent rather than a reactive chat tool. And the standalone Claude design site now displays a notice that it will close on December 14th, so anyone using it should migrate work before that date.
Frequently Asked Questions
What is the cheapest Claude model available in Claude Code now?
Haiku 5.5 is the cheapest, priced at 10 cents per million input tokens and 50 cents per million output tokens for prompts under 100,000 tokens, rising to 50 cents and $2.50 beyond that.
Does Ultra Think cost more to use?
Typing Ultra Think raises the reasoning effort for that turn, which generally means more computation and higher token usage, but Anthropic added an environment variable to disable the behavior if you want to avoid the extra cost.
Can I turn off idle compaction?
Yes. It’s controlled through a dedicated row in the /config menu, which is useful if you need Claude to retain full conversational detail rather than a summarized version after long idle periods.
Is the standalone Claude design site still usable?
It is for now, but Anthropic has announced it will close on December 14th, so any work stored there should be moved before that date.
What is Claude Duet?
It’s a separate app Anthropic launched October 21st as a research preview for Pro and Max subscribers, built around Claude acting as a proactive agent rather than responding only when prompted.
