Claude Code

Claude Code 2.1.280 adds Opus 5.5 and makes Opus the default on Pro

3 min read AI-generated

Version 2.1.279 does not exist; the jump goes straight from 2.1.278 to 2.1.280. And an effort level you saved earlier no longer applies to newly released models.

Featured image for "Claude Code 2.1.280 adds Opus 5.5 and makes Opus the default on Pro"

Claude Code 2.1.280 hit the npm registry yesterday at 15:44 UTC, shortly before Anthropic’s post on Claude Opus 5.5. It is the biggest release in weeks, and a good share of it is housekeeping.

Side note: 2.1.279 does not exist. Neither the registry nor the changelog knows the number; the jump goes straight from 2.1.278 to 2.1.280.

Opus 5.5 arrives with a million tokens of context

The first changelog entry has the specs: claude-opus-5-5, now the default Opus model, 1M context, $4 and $20 per million tokens, $0.20 for cache reads.

The second point matters more to a lot of people. On Pro and Team Standard, the default model switched from Sonnet to Opus. All paid plans now start the same way, since Max, Team Premium and Enterprise already did. If you deliberately started on Sonnet, you now have to pick it yourself.

A third change hangs off that one. An effort level you saved before /effort became per-model no longer applies to newly released models. Opus 5.5 starts at its own default until you choose a level.

The fix that tells me the most sits further down the list. Until now, a write through a symlinked path was judged by how it was spelled inside the project, not by where it actually landed. acceptEdits, allow rules and auto mode could wave through a write that ended up outside the tree. The prompt now names the real destination.

Two more corrections hit auto mode directly. When a safety check declined to review an action, Claude Code kept retrying it; the action is now denied once, with a note that retrying won’t help. And when a check gave no answer at all, denials ran in a loop with no pause. Retries now back off, and after ten in a row the turn stops with a message.

A long list of small things

The rest is unglamorous and touches things you trip over daily. A stray n used to close dialogs and a stray y to confirm them; that is now Enter and Escape only. Text fields lost typed letters to keybindings. Ctrl+C or Ctrl+D pressed twice quit Claude Code in most dialogs instead of closing the dialog. Home and End did nothing in /config and in selection lists.

Then there is a pile of crashes and error loops: sessions failing on every turn with an API error, damaged transcripts on resume, a hang when a settings file is replaced by a named pipe mid-read. VS Code gains dialogs for status, sandbox, Claude in Chrome and skills, plus a conversation export.

The most interesting line is the shortest one

The symlink fix is not cosmetic. Auto mode decides without asking, and the decision rested on a path that did not lead where it appeared to lead. That is exactly the kind of gap that vanishes in a release with sixty entries.

What Anthropic showed the same day, a model that runs 18 hours unattended, makes fixes like that more important, not less. The longer an agent works without a human watching, the more depends on the rule that fences it in checking the right path.

Sources

Claude CodeReleaseOpus