Claude Code updates in October 2026: the 8 changes that hit your wallet and your workflow

The Claude Code updates of October 2026 came in 22 releases, from 2.1.287 (October 1) to 2.1.310 (October 23): Haiku 5.5 as the default Haiku, idle compaction before the cache expires, cost warnings in /usage, and fork subagents on by default outside the terminal.

Almost nobody read the whole changelog. Fair enough: it's a wall of "Fixed" and "Added" that runs past 9,000 lines.

The problem is that buried in that wall are changes that hit how much you pay and when your 5-hour limit runs out. And changes that alter how the agent behaves without you asking for anything.

This post is the filtered recap of what changed in Claude Code in October 2026. It's not a changelog dump. These are the 8 changes that matter if you use Claude Code every day, with the version number for each one (so you can check whether you're on it) and what to do about it.

TL;DR

• What it is: a recap of Claude Code versions 2.1.287 to 2.1.310 (October 2026), filtered by cost and workflow.
• Wallet: Haiku 5.5 at $0.10/$0.50 per MTok, idle compaction before the 1-hour cache expires, new /usage warnings, and a limit warning when resuming a conversation.
• Workflow: fast mode without the preview label on Opus 5.5, /auto-mode-setup, fork subagents on by default outside the terminal, Monitor tool on Bedrock/Vertex/Foundry.
• On the way out: standalone Claude Design closes by 12/14/2026.
• Source: official CHANGELOG.md and the Haiku 5.5 page.

Why following the Claude Code changelog matters (and why nobody reads it)

Claude Code updates itself. You open the terminal on a Monday and it's already three versions ahead.

That's great for bug fixes. It's terrible for behavior changes. When a release turns on fork subagents by default in claude -p, or starts compacting your conversation while you're at lunch, your pipeline changes without a single commit from you.

And in October the dominant theme was one thing: context cost. Half of the relevant changes revolve around prompt caching, the 5-hour limit and reprocessed tokens. If you've read our guide on saving tokens in Claude Code by not breaking the cache, you'll recognize the pattern: Anthropic is pushing into the tool the good practices that used to be your job.

Tracking 20+ releases a month on your own doesn't scale. That's why at Clã Beer and Code we filter this kind of change every week, live, looking at what it does to the harness, the agents and the bill of people running AI in production.

Here are the 8.

1. Haiku 5.5 arrived and became the default Haiku (2.1.293)

Claude Haiku 5.5 launched on October 7 and, the same week, Claude Code 2.1.293 added claude-haiku-5-5 as the default Haiku on the Anthropic API.

The numbers that matter:

The detail that changes the math: the $0.10/$0.50 prices apply to prompts up to 100K tokens. Above that, it goes to $0.50/$2.50. The changelog lists a 1M context window.

What to do with it: Haiku 5.5 is not your coding model. 39.2% on Terminal-Bench against Sonnet 5.5's 70.6% makes that clear. It's the model for short, high-volume work: search subagents, summaries, classification, the hook that calls a model by alias (haiku). If you have an agent hook or subagent pointing at haiku, it now resolves to a model with a list price ten times lower than Haiku 4.5 (Anthropic itself talks about roughly 75% savings on average, because the price goes up above 100K tokens). Check that quality held up on your cases before celebrating.

2. Idle compaction: Claude Code compacts before the cache expires (2.1.299)

This is the most important one of the month if you work in long sessions.

The classic scenario: you're in a 300K-token conversation, you leave for a one-hour meeting, come back and type "ok, keep going". The 1-hour prompt cache has already expired. Your next message reprocesses all 300K. No cache. On your quota.

2.1.299 fixed this in the terminal: a 200K+ token conversation left idle for about 54 minutes is compacted before its one-hour cache expires. When you come back, your next message doesn't reprocess everything, it reprocesses the summary.

You can tune the trigger or turn it off:

# muda o mínimo de 200K tokens para disparar a idle compaction
export CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS=400000

{
  "idleCompaction": false
}

The second block goes in settings.json. Both need v2.1.290 or later. In 2.1.302 an Idle compaction row was also added to /config to toggle it without editing a file, and the post-compaction notice now tells you how to turn it off.

When to turn it off: if you depend on fine detail from the history (a full stack trace, a decision from 100 messages ago), compaction loses information. In that case, either turn it off or raise the minimum. For everything else, leave it on: it's money you were burning for nothing.

3. /usage now points at the problem and fixes it with one key (2.1.299 and 2.1.305)

/usage stopped being just a counter.

In 2.1.299 it got warnings for four consumption villains: max effort, workflows, always-on context and large tool results.

In 2.1.305 each warning became a one-line fix that names its source, and pressing f applies the fixes. The same suggestions show up in VS Code and in /doctor.

In practice, /usage became a cost linter. If your CLAUDE.md loads 40K tokens into every session, or an MCP server returns giant JSON on every call, the tool now tells you exactly where it's coming from.

What to do: run /usage today and read the warnings before pressing f. The auto-fix is convenient, but you need to know what it turned off.

4. A warning before resuming a conversation that will eat your 5-hour limit (2.1.304)

Same problem as item 2, from the other end.

2.1.304 added, on Pro and Max plans, a warning when resuming an inactive conversation would use a noticeable share of your 5-hour usage limit. You choose: resume anyway, or start a new conversation and Claude can look up the previous one if you refer to it.

It's the kind of detail that explains a lot of "my 5-hour quota ran out in half an hour" reports. Resuming a fat session without cache is expensive, and until now nobody warned you. If you want to understand how the limit works and what does or doesn't reset it, we broke it down in the post about the Claude usage limit reset.

What to do: when in doubt, take the new conversation. Starting clean with a good summary is almost always cheaper than dragging 300K cold tokens along.

5. Fast mode loses the preview label on Opus 5.5 (2.1.304)

2.1.304 removed the "(research preview)" label from the /fast dialog title on Opus 5.5.

Looks cosmetic. It isn't. It signals that fast mode became an official feature, and it has its own math: same model, more output tokens per second, higher price. We did the full math, with numbers and when it's worth it, in the post about Claude Code fast mode.

Two related changes this month: 2.1.299 added /effort and /fast in agent view to set what new sessions start with, and 2.1.300 fixed a VS Code bug where fast mode stayed on after usage credits ran out, so every request was refused and re-sent. Now it turns off and shows an error.

What to do: if you use VS Code with fast mode, update. And don't make fast mode your global default without doing the math.

6. Auto mode got a guided setup and a configurable model (2.1.302 and 2.1.304)

Auto Mode is the classifier that approves tool calls on your behalf. In October it got two levers:

• /auto-mode-setup (2.1.304): a guided setup that reads your project and recent sessions to teach auto mode about your environment. Also available on Bedrock, Vertex, Foundry, the Claude apps gateway and telemetry-disabled sessions.
• CLAUDE_CODE_AUTO_MODE_MODEL (2.1.302): picks which model the client-side classifier uses. Mind the detail: it doesn't change the model for server-side auto mode.

export CLAUDE_CODE_AUTO_MODE_MODEL=claude-haiku-5-5

Tying it back to item 1: a permission classifier is exactly the kind of short, repetitive task where a cheap model makes sense. But it's also a security layer. Swapping the classifier model to save money is a risk decision, not a cost decision. Test first with commands you know should be blocked.

7. Fork subagents on by default outside the terminal (2.1.302)

This is the one that can change your pipeline's behavior without warning.

Fork subagents (subagent_type: "fork") are subagents that inherit the full conversation context. In the terminal they were already on. 2.1.302 turned them on by default in claude -p, the Agent SDK, the desktop app and VS Code too.

If you orchestrate Claude Code from code, like in our example of headless Claude Code orchestrated by PHP, the agent can now fork into parallel tasks that carry the entire history. That can speed things up, and it can multiply token usage.

To go back to the old behavior:

export CLAUDE_CODE_FORK_SUBAGENT=0

claude mcp serve never offers forks. And in 2.1.307, forks launched together are now told which launch is their own, which cuts down on one fork trying to do another's work.

What to do: if you run headless batches, compare the token usage of one run before and after 2.1.302. If it went up with no quality gain, turn it off.

8. The Monitor tool reached Bedrock, Vertex and Foundry (2.1.310)

Last release of the month. 2.1.310 brought the Monitor tool and plugin monitors to Bedrock, Vertex and Foundry, and to sessions with telemetry or nonessential traffic disabled.

The Monitor tool is what lets the agent watch a background process (a build, a server, a log) and react when something happens, instead of sleeping and polling. It came with a guardrail: long foreground sleeps are now blocked in those environments too.

For companies running Claude Code through a cloud provider for compliance reasons, this closes a feature gap against people using Anthropic's API directly.

What to do: if your script or skill did sleep 300 waiting for something to finish, it will break in those environments. Replace it with a conditional wait or with the Monitor.

On the way out: standalone Claude Design closes on 12/14

Not a new feature, a shutdown notice. 2.1.302 added a notice to /design-sync and /design-login: standalone Claude Design (claude.ai/design) closes by December 14, 2026.

If you use the prototype-to-code flow we showed in the Claude Design guide for devs, export what you have there before the date.

Other Claude Code updates in October

They didn't make the 8, but they're worth a line:

• CLAUDE_CODE_DISABLE_FILE_MENTIONS (2.1.299): @ mentions pass only the file path and Claude reads it with its own tools, instead of injecting the whole content into the prompt.
• A more honest /usage (2.1.304): when the rate limit on usage data kicks in, it tells you when it ends instead of "try again in a moment".
• If you want to get the most out of the extension layer that evolved most this year, the tutorial on how to use skills in Claude Code covers structure, frontmatter and when a skill beats MCP.

Limitations and things to watch

• The changelog has no date per version. The dates in this post come from npm releases. The version you get depends on when auto-update ran.
• Default changes bite the hardest. Fork subagents and idle compaction change behavior without you asking. If you run pipelines in production, you need a pinned version or regression tests.
• The /usage auto-fix is a shortcut, not a diagnosis. It may turn off something you wanted on.
• A small model's benchmark is not your use case. The Haiku 5.5 numbers are Anthropic's. Run it on your own dataset before swapping a production subagent.

FAQ

How do I check which Claude Code version I'm on? Run claude --version in the terminal. To force a Claude Code update right away, claude update.

Does idle compaction work in VS Code and the desktop app? The 2.1.299 changelog says "idle compaction in the terminal". For the other clients, there's no entry in October.

Does Haiku 5.5 replace Sonnet in Claude Code? Not for heavy coding work. It comes in as the cheap model for subagents, hooks and classification. For coding, Sonnet 5.5 stays well ahead: 70.6% vs 39.2% on Terminal-Bench 4.0.

How do I turn off fork subagents in claude -p? CLAUDE_CODE_FORK_SUBAGENT=0 in the environment.

Where do I follow the official changelog? In the CHANGELOG.md of the anthropics/claude-code repo. Or wait for next month's recap here.

Conclusion

October was the month Claude Code started managing your cache for you. Idle compaction, the resume warning, /usage with fixes and Haiku 5.5 at a list price ten times lower all point the same way: context cost became Anthropic's product problem, not just yours.

But the tool only covers the obvious. Knowing when compaction loses critical information, when a fork subagent is waste, and when swapping the classifier model becomes a security risk is still engineering.

Update, run /usage, check your headless pipelines. And come back at the end of November for the next recap.