Claude Code 2.1.311 makes ultrathink raise effort, and 2.1.312 renames your limit

What happened

Anthropic published Claude Code 2.1.311 on October 26 and 2.1.312 on October 27, 2026.

Ultrathink raises effort for one turn

In 2.1.311, typing ultrathink in the terminal prompt can raise that one turn to xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic's API. Your session's effort setting stays where it was. When the keyword can work, it shows in rainbow colors. Typed while Claude is already busy, it raises the rest of that turn.

Ultrathink stays plain text when the word sits inside quotes, brackets, a path or a command; the prompt starts with /; your CLAUDE_CODE_EFFORT_LEVEL or maxEffortLevel is set below xhigh; or you're on Amazon Bedrock.

"Think hard" still passes through as ordinary text.

Usage limits

• No more "Usage credits required for 1M context". A standard-context session that ran past 200K tokens used to stop with that error. Since 2.1.311 Claude Code compacts the conversation and carries on.
• /usage shows cache misses. Large prompt-cache misses now appear under "What's contributing to your limits usage?", so you can see when switching models or effort mid-session cost you.
• "5-hour limit", not "session limit". In 2.1.312 the messages call the rolling window by its length, and in VS Code the reset shows a clock time, such as "Resets at 3:10 PM".

Your CLAUDE.md is safer

2.1.312 fixes Claude overwriting a very large CLAUDE.md, rules file or @-included file in a subfolder after reading only its start. The Write tool now tells Claude to read the whole file or edit it in place.

Smaller fixes

• Prompt and agent hooks now run at most 24 at a time per session (2.1.312).
• SubagentStop hooks fired about twice a minute for every running subagent; they now fire once, when it finishes (2.1.311).
• Claude Code no longer retries requests the API marked as not retryable, which could hold a turn for minutes (2.1.311).
• Folder-scoped skills survive a cd (2.1.312).

What it means for you

Run claude update, then check that claude --version shows 2.1.312. That matters most if your project keeps long CLAUDE.md or rules files.

For the project setup that keeps sessions short, see how to set up checks, subagents and permissions.

Try it

ultrathink. This bug has survived two fixes already. Before you change anything:
1) list every place the failing value is written or read,
2) give three hypotheses for the cause, each with the one check that would rule it out,
3) run those checks and report what each showed.
Only then propose a fix, and say which test proves it.

The bug: [what happens, what you expected, how to reproduce]

I'm going to paste the output of /usage and a short list of what I did in this session. Tell me:
1) which part of my work used the most of my limit,
2) whether model or effort switches mid-session caused cache misses, and where,
3) three changes to how I work tomorrow that would use less, in order of impact.
Keep it to what the numbers show. If they can't answer a question, say so.

/usage output:
[paste]
What I did:
[list]

Sources

• Anthropic, Claude Code changelog, versions 2.1.311 and 2.1.312, checked October 28, 2026
• npm, @anthropic-ai/claude-code release history, checked October 28, 2026
• Claude Code docs, Model configuration: use ultrathink for one-off deep reasoning, checked October 28, 2026