Source: raw/anthropic-watch-claude-code-tag-v2-1-{280,281,282,283}.md — anthropics/claude-code GitHub release notes, published 2026-09-22 to 2026-09-25 — plus anthropics/skills commit 33375500 (2026-09-24). The watch-feed copy of v2.1.281 is truncated (“8313 of 28308 chars dropped”); its remainder was read from https://github.com/anthropics/claude-code/releases/tag/v2.1.281 via gh release view on 2026-09-29. Also ai-research/claude-code-docs-changelog-2026-09-25.md (the code.claude.com changelog, v2.1.280–283, fetched 2026-09-25). No official W39 digest existed yet: code.claude.com/docs/en/whats-new/2026-w39 returned HTTP 404 on 2026-09-25.

Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases · Date: 2026-09-22 → 2026-09-25

Four large releases (v2.1.280 through v2.1.283). The headline is Claude Opus 5.5 becoming Claude Code’s default Opus model on September 22, together with Pro and Team Standard plans switching their default from Sonnet to Opus; full coverage is in Claude Opus 5.5. Admins get availableModelsMatch: "exact" and deniedModels to hold new models back until they approve them, and /doctor prompt-audit checks CLAUDE.md files, skills, agents and commands for prompting patterns written for older models. Continues Week 38.

Key Takeaways

  • Claude Opus 5.5 (claude-opus-5-5) is the default Opus model — 1M context, 20 per Mtok (20% below Opus 5’s 25), $0.20/Mtok cache reads (v2.1.280, 2026-09-22). See Claude Opus 5.5.
  • Pro and Team Standard plans now default to Opus, not Sonnet, “matching Max, Team Premium, and Enterprise” (v2.1.280). Every plan named in the release notes now starts on Opus.
  • New models start at their default effort. An effort level saved before /effort became per-model no longer applies to newly released models such as Opus 5.5 (v2.1.280). Opus 4.7, Opus 4.8 and Fable 5 stop holding their launch-default effort over /effort in -p/Agent SDK, project/managed/--settings effortLevel, or a per-model level (v2.1.280). See Model vs Effort.
  • Admins can pin exact model versions. With availableModelsMatch: "exact", an availableModels entry allows only the version it names, so new releases stay blocked until listed; deniedModels blocks specific models even when availableModels allows them (v2.1.283).
  • /doctor prompt-audit (also /checkup prompt-audit) audits CLAUDE.md files, skills, agents and commands for prompting patterns written for older models; stale paths, stale commands and contradicting instruction files lead the report (v2.1.283). See CLAUDE.md Primer.
  • Agent-type hooks no longer run on PermissionRequest, “since its answer could never allow or deny the request”; an error points to command or http hooks (v2.1.280). See Hooks.
  • Namespace reservation, half-reverted. v2.1.282 stopped loading skill folders, command files and workflows named in the anthropic-skills or claude-ai namespace; v2.1.283 reverted the claude-ai half. The anthropic-skills reservation stands. See Agent Skills.
  • Auto mode on third-party providers or with telemetry off now starts by default when no permission mode is configured (v2.1.283) — extended to every plan and provider the next week in Week 40. See Auto Mode.
  • "attribution": false in settings.json hides all commit and PR attribution; older CLIs skip a settings file holding it, so keep the object form in files shared across versions (v2.1.281).

Release table

VersionDate (UTC)Shape
v2.1.279—No GitHub release page
v2.1.2802026-09-22Claude Opus 5.5; Pro/Team Standard default → Opus; PermissionRequest agent hooks removed
v2.1.2812026-09-23"attribution": false, /insights auto-mode estimate, MCP URL-mode elicitation, proxy stream fixes
v2.1.2822026-09-24maxProseWidth, namespace reservation, OTel variables ignored in project settings
v2.1.2832026-09-25availableModelsMatch/deniedModels, /doctor prompt-audit, auto mode default on 3P

Models and defaults

  • Claude Opus 5.5 added and made the default Opus model (v2.1.280).
  • Pro and Team Standard default model: Sonnet → Opus (v2.1.280). The code.claude.com changelog snapshot of 2026-09-25 lists this change, and the effort change below, under v2.1.282 as well; the v2.1.282 GitHub release notes contain neither.
  • Pre-per-model saved effort no longer carries to new models; Opus 4.7/4.8 and Fable 5 stop holding launch-default effort (v2.1.280).
  • availableModelsMatch: "exact" and deniedModels (v2.1.283).
  • The /model picker’s Opus row and Default model name drop “(1M context)” where Opus already has a 1M window; the window is unchanged (v2.1.283).
  • Alt+T and /config no longer offer to turn thinking off on models that can’t; thinking stays on with a one-line reason (v2.1.281).
  • Fable fixes: an unanswered Fable usage-credits prompt no longer switches models in SDK-hosted sessions such as Claude Desktop (v2.1.282); /model with a full Fable ID opens the usage-credits prompt instead of stopping at an API error (v2.1.282); the weekly Fable limit appears in /usage with telemetry off (v2.1.283).
  • A failed turn (“Effort ‘xhigh’ isn’t available with thinking turned off”) after a safety-related model switch — fixed (v2.1.282).
  • Compaction refused by the summarizer now retries on a fallback model (v2.1.282).
  • Dynamic workflows started during a model fallback retry the configured model instead of running every agent on the fallback (v2.1.283).
  • /model accepting Sonnet 4.6 or Sonnet 5 with [1m] on dated or -v1:0 IDs fixed; the Haiku row respects ANTHROPIC_DEFAULT_HAIKU_MODEL; DISABLE_PROMPT_CACHING_HAIKU works when Haiku is the main model (v2.1.283).

New commands and features

  • /doctor prompt-audit (v2.1.283).
  • "attribution": false (v2.1.281).
  • /insights auto-mode recommendation estimating how many permission prompts auto mode could have handled in recent sessions (v2.1.281).
  • MCP URL-mode elicitation on 2026-07-28 protocol connections: servers can ask Claude Code to open a browser-based flow (v2.1.281). See MCP 2026-07-28 Spec.
  • maxProseWidth caps the width of Claude’s prose in wide terminals while tables and code keep full width (v2.1.282).
  • Send-now (ctrl+enter / ctrl+x ctrl+s) moves running tools to the background instead of cancelling the turn (v2.1.281).
  • --agents accepts a path to a JSON file (with -p) and an empty prompt (v2.1.281).
  • /batch runs where a WorktreeCreate hook provides the agent worktrees, not only inside a git repository (v2.1.281).
  • /deep-research reliability improved on long research briefs (v2.1.281).
  • Artifacts: session artifact links collapse into one footer pill (⧉ name or ⧉ N) that opens /artifacts; the Artifact tool may load scripts from unpkg.com (v2.1.281); auto-armed artifact watches end after 3.5 hours with no activity (v2.1.283). See Artifacts.
  • Mouse support, page keys and scrollbars in more fullscreen lists and pickers, including /skills, /mcp, /plugin Installed, /workflows, /artifacts and /feedback (v2.1.280–283).
  • MCP tool images are also saved to a file so Bash and Read can open them (v2.1.283).
  • The compaction spinner counts the summary’s tokens as they stream (v2.1.283).

Settings and environment variables

  • CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH changes the 2,048-character cap on MCP tool descriptions and server instructions (v2.1.280). See MCP.
  • CLAUDE_CODE_AUTO_MODE_SERVER now also applies on direct Anthropic API connections: 0 opts out of the server-side classifier (the local classifier then counts toward usage), 1 opts in (v2.1.281); the server-side classifier is the default on a direct API connection with telemetry off (v2.1.282).
  • CLAUDE_CODE_DISABLE_DANGEROUS_RM_TIMEOUT=1 — otherwise the dangerous-rm prompt under --dangerously-skip-permissions or auto mode waits 2 minutes, then denies with a rewrite hint (v2.1.281).
  • CLAUDE_CODE_DISABLE_SUBSTITUTION_RM_PROMPT=1 opts out of the new prompt for rm -rf "$(pwd)"-style targets (v2.1.281).
  • CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 gains x-claude-code-prompt-id, so LLM gateways can group the requests that serve one user prompt (v2.1.283). See Claude Apps Gateway.
  • OpenTelemetry: OTEL_LOG_TOOL_CONTENT=1 includes MCP tool, WebFetch and WebSearch outputs in the tool.output span event (v2.1.283); hook_execution_complete gains output sizes (v2.1.280); project and local settings ignore OTel variables that turn on export, set the endpoint or capture content (CLAUDE_CODE_ENABLE_TELEMETRY, OTEL_LOG_*), with a startup notice listing them (v2.1.282).
  • allowClaudeInChromeWithManagedMcp lets claude --chrome run alongside an exclusive managed-mcp.json (v2.1.282).
  • sandbox.excludedCommands from project and local settings is ignored when managed settings set allowUnsandboxedCommands: false or allowManagedDomainsOnly: true (v2.1.282).
  • --system-prompt and --append-system-prompt accept their text and -file forms together; the file’s text comes first (v2.1.283).
  • Self-hosted runners pass system prompts as private files; wrappers or command hooks that append --system-prompt must switch to --system-prompt-file / --append-system-prompt-file (v2.1.281).
  • Gateway: assume_role on Bedrock upstreams for cross-account IAM access — the gateway assumes the role through STS, optionally one session per developer — and guardrail: {id, version}, which applies an Amazon Bedrock guardrail to every request through the upstream (set on all Bedrock upstreams or none), plus telemetry.resource_attributes (v2.1.281); store.readiness_grace_seconds (v2.1.282); load_test_mode and a mantle upstream provider for Bedrock’s Mantle endpoint (v2.1.283).

Hooks, MCP, plugins, skills and subagents

  • PermissionRequest: agent-type hooks removed (v2.1.280).
  • mcp_tool hooks on blocking events (PreToolUse and similar) wait for their MCP server to connect, up to the MCP connect timeout, instead of being skipped (v2.1.281).
  • The UserPromptSubmit hook timeout notice names the hook (v2.1.280); the /hooks detail screen says what kind of hook it is and where to change it (v2.1.281).
  • claude plugin validate checks MCP servers (dropped .mcp.json entries, undeclared ${user_config.*}, insecure URLs) and warns on unquoted ${CLAUDE_PLUGIN_ROOT} in shell-form hooks (v2.1.281); it also checks outputStyles, themes, monitors and lspServers paths (v2.1.283). claude plugin eval requires git 2.31+ when git is installed (v2.1.283). See Plugins and Marketplaces.
  • Skills in ~/.claude/skills/ are no longer moved to trash when a manifest.json in that folder listed them (v2.1.280); claude.ai-synced skills show by their short name when nothing else uses it (v2.1.281).
  • anthropic-skills / claude-ai namespace reservation (v2.1.282): skill folders, command files and workflows so named stop loading; MCP servers so named list no skills or prompts; Skill(anthropic-skills:*) and Skill(claude-ai:*) allow rules cover only synced skills. v2.1.283 reverted the claude-ai half. Skill(anthropic-skills:<name>) deny rules also block the skill when Claude Desktop delivers it as a plugin (v2.1.283).
  • Repo, user and --add-dir skills and commands can no longer pre-approve their own tools via allowed-tools under managed allowManagedPermissionRulesOnly (v2.1.282).
  • Subagents: background subagents can use the LSP tool (v2.1.280); resumed fork subagents keep their original tool list, preserving the cache (v2.1.280). See Subagents.
  • MCP: the same server no longer connects twice when a plugin or connector spells its URL differently (v2.1.281); a brief 404 from a stateless remote server no longer leaves it unusable (v2.1.283); /context counts MCP server instructions (v2.1.283); progress notifications are kept after a long tool call moves to the background (v2.1.283); MCP resource lists skip MCP Apps UI resources (v2.1.281).
  • The leftover “(removed)” /agents entry is gone from the menu; typing /agents still explains where the wizard went (v2.1.281).

Breaking and behaviour changes

  • ctrl+l / cmd+k clearing the transcript view (added in v2.1.260) is reverted; they redraw the screen again (v2.1.280).
  • A stray y or n no longer confirms or closes dialogs — Enter and Esc do; bind confirm:yes / confirm:no to restore (v2.1.280).
  • Plugin marketplaces whose name imitates a reserved marketplace name are refused, and stop loading if already added (v2.1.280).
  • Interactive sessions on third-party providers or with telemetry off start in auto mode when no permission mode is configured; permissions.defaultMode still overrides (v2.1.283).
  • Where auto mode’s classifier runs server-side, read-only and sandboxed shell commands also wait for review and are blocked when flagged (v2.1.281).
  • Queued messages show above the spinner (v2.1.281).
  • The /ultrareview launch dialog warns that reviewing a local branch may upload uncommitted changes to tracked files (v2.1.283). See Ultrareview.
  • New routine schedules default to a few minutes past the hour, since routines set exactly on the hour can start late (v2.1.283).

The claude-api skill: refusal billing (skills 33375500, 2026-09-24)

The claude-api skill now links each refusal-billing sentence to the “How refusals are billed” docs section instead of restating the rules, and corrects two claims: a refusal before any output counts against rate limits, and a mid-stream refusal bills input tokens as well as the output already streamed. Claude Code v2.1.282 shipped the same update to the bundled skill.

Notable fixes

  • Security: writes through a symlinked path judged by their in-tree spelling, so acceptEdits, allow rules and auto mode could approve a write landing outside (v2.1.280); rm -rf "$(pwd)" running unprompted in auto and skip-permissions mode (v2.1.281); claude --bg running project hooks in an untrusted directory (v2.1.281); --setting-sources not forwarded to teammates and spawned sessions (v2.1.281); PowerShell cmd /c rd deleting drive roots and the home folder (v2.1.283).
  • Proxy and gateway streams: responses cut short by a clean stream close shown as complete, tool calls running twice on duplicated events, “Content block not found” on a dropped event (v2.1.281).
  • Prompt cache and reasoning preservation across resume and slash commands (v2.1.281, v2.1.282), including resumed sessions re-sending earlier turns in a changed form, which could make the API drop the conversation’s prior reasoning.
  • Auto mode retrying an action over and over when a safety check declined to review it; the action is now denied once (v2.1.280).
  • SDK sessions losing a deferred tool call or finished tool result when a turn ended early (v2.1.283). See Agent SDK.
  • Stdio MCP servers left running when the session ended while they were still starting (v2.1.283).
  • Bash and PowerShell hiding a full disk quota behind “Exit code 1” (v2.1.282).
  • “Invalid data in redacted_thinking block” failing every turn: thinking blocks are dropped and the turn retried once (v2.1.282).
  • The advisor’s advisor_20260301 400 behind proxies now retries without the advisor (v2.1.280). See Advisor Tool.
  • Ctrl+C / Ctrl+D pressed twice in dialogs quitting Claude Code (v2.1.280, v2.1.281).

Try It

  • Pro and Team Standard users: check /model — your default is now Opus; pick Sonnet there if you preferred it.
  • Admins who roll out models deliberately: set availableModelsMatch: "exact" and list the versions you have approved.
  • Run /doctor prompt-audit on a long-lived repo; stale paths and contradicting instruction files come first in the report.
  • If you wrote an agent-type PermissionRequest hook, rewrite it as a command or http hook.
  • If you put OTel export settings in a project settings file, move them to user or managed settings; project files no longer apply them.
  • Want no attribution lines at all? Add "attribution": false to ~/.claude/settings.json; use the object form in settings files that older CLI versions also read.
  • Bedrock gateway operators with several AWS accounts: set assume_role on each cross-account upstream instead of provisioning Bedrock access in every account that runs a gateway.
  • Verbose MCP servers: check /context (which counts MCP server instructions from v2.1.283) and set CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH if their descriptions take a large share.

Open Questions

  • v2.1.279 has no GitHub release page, and the code.claude.com changelog skips it too, as it skipped v2.1.230 and v2.1.262.
  • Why does the code.claude.com changelog list the Pro/Team Standard default change under both v2.1.280 and v2.1.282 when the GitHub notes list it once? A staged rollout or a changelog duplication; neither source says.
  • Why was the claude-ai namespace reservation reverted after one day while anthropic-skills stayed? No reason is given.
  • What prompting patterns does /doctor prompt-audit flag as written “for older models”? The release note names only stale paths, stale commands, contradicting files and kept thinking keywords.
  • What form does the /doctor prompt-audit report take (terminal only, or a file)? Neither source says.