Claude Code 2.1.280: Opus 5.5 as the default everywhere, a symlink permission gap closed

Claude Code 2.1.280, released on September 22, 2026, packs about a hundred lines of changes, but only three actually matter. The first refreshes the default model and extends it to plans that didn't have it yet. The second closes a real permission gap: a write through a symlink could land outside the intended folder without tripping the safeguards meant to catch it. The third is a new environment variable for anyone using MCP servers with long tool descriptions. Everything else, roughly a hundred interface and reliability fixes, is summarized in a quick list further down.

Opus 5.5 is now the default, including on Pro and Team Standard

Claude Opus 5.5 (claude-opus-5-5) is the new default Opus model. The models overview recommends it directly: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads." It ships with a 1M token context window, 128K tokens of max output, and costs $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million, 5% of the input price. Its default effort level is medium, lower than most other Opus models, which default to high.

Opus 5.5 just shipped, on September 22, 2026, and directly replaces Opus 5 as Claude Code's default Opus model from 2.1.280 on: per the model configuration docs, default now resolves to Opus 5.5 on Pro, Max, Team, Enterprise, the Anthropic API, Claude Platform on AWS, Amazon Bedrock, and Google Cloud, while Microsoft Foundry stays on Sonnet 4.5.

The second change in the same release: the default model switches from Sonnet to Opus on Pro and Team Standard plans, bringing them in line with Max, Team Premium, and Enterprise, which already defaulted to Opus. If you're on either of those two plans, your next session now starts directly on Opus 5.5 with nothing to change in /model.

One nuance affects effort levels set the old way. The effort adjustment docs spell it out: "A top-level effortLevel in your user settings file doesn't count for Opus 5.5 [...] while Opus 5.5 and models released after it start at their own default until you choose a level for them with /effort or the /model picker." A top-level effortLevel in your user settings, set before /effort started saving a level per model still applies to Opus 5, Fable 5.1, and earlier models, but not to Opus 5.5, which starts at medium until you explicitly pick a level for it.

Security: a symlinked write could land outside the approved folder

The fixed bug is specific. The permissions docs explain how a symlink is normally handled: "When Claude accesses a symlink, permission rules check two paths: the symlink itself and the file it resolves to," and an allow rule only applies when both the link and its target match. That principle exists precisely to stop a symlink pointing outside an approved folder from riding on an allow rule meant for that folder.

The 2.1.280 bug bypassed that principle on the write side: an action was judged by how the path was spelled in the file tree, not by where it actually landed once the symlink resolved. As a result, acceptEdits, an allow rule, or auto mode could approve a write that looked like it stayed inside the working directory while it actually landed, via a symlink, on a path that was never authorized. That's fixed now: the check runs against the real destination, not the path's surface spelling, and the permission prompt now names where the write actually lands. If your project has symlinks and you rely on acceptEdits or auto mode to skip approving every write, this release closes a door you probably didn't know was open.

MCP: a description length cap you can finally adjust

Claude Code caps every MCP tool description and server instruction at 2,048 characters before sending it to the model. 2.1.280 adds CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH to change that cap for every MCP server in the session at once. Useful if you connect an MCP server whose detailed tool descriptions were getting truncated, at the cost of a few more tokens sent on every turn.

Everything else, in brief

• Auto mode: an action a safety check declined to review no longer gets retried in a loop, it's now denied once with a note that retrying won't help. Conversely, repeated denials with no answer from a safety check now back off gradually, and the turn stops with a message after ten denials in a row.
• Plugin marketplaces whose name imitates a reserved marketplace name are now refused when added, and stop loading if they were already added.
• Stray keystrokes: a stray y or n no longer closes or confirms a dialog on your behalf; only Enter and Esc do (y/n can still be rebound in keybindings.json).
• Windows terminal: the prompt line no longer stays scrambled after invisible-character cleanup on Enter; the screen now repaints so you review the exact text that will be sent.
• claude --bg: background sessions can run git, hooks, plugins, and other helper programs again when an environment variable handed to the session contains a NUL character.
• VS Code: new typed /status, /sandbox, /chrome, /export, and /plan commands; the Sandbox and Chrome dialogs also open from the panel menu.
• Claude Code on the web: the admin routines toggle moved under Admin settings → Capabilities → Remote sessions.

The takeaway

If you're on Pro or Team Standard, your next session runs on Opus 5.5 with no action needed, at a lower default effort (medium) than you might expect from Opus: if you were counting on high for this model, go back through /effort. The permission fix is worth taking seriously if your project uses symlinks: it closed a real gap between what acceptEdits or auto mode believed they were approving and what actually happened on disk. The rest of this release is interface maintenance, worth skimming only if one of the points above applies to you directly.