Source: ai-research/claude-code-release-v2-1-{252,257,258,259,260,261,263}.md — anthropics/claude-code GitHub release notes fetched with gh release view on 2026-09-29 — plus anthropics/skills commits 53048666 (claude-api skill, 2026-09-01) and 41bbe19d (frontend-design skill, 2026-09-03). Also ai-research/claude-code-docs-whats-new-w36-2026-09-18.md, the official W36 digest fetched from code.claude.com/docs/en/whats-new/2026-w36 on 2026-09-18. The code.claude.com changelog snapshot of the same day (ai-research/claude-code-docs-changelog-2026-09-18.md) starts at v2.1.268, so this week’s releases rest on the GitHub notes.

Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases · Official digest: https://code.claude.com/docs/en/whats-new/2026-w36 · Date: 2026-08-31 → 2026-09-06

Seven releases this week (v2.1.252, 257, 258, 259, 260, 261, 263). The headline is Claude Fable 5.1 becoming Claude Code’s default Fable model on September 1, with the bundled claude-api skill naming Claude Mythos 5.1 the same day; full coverage is in Claude Fable 5.1 + Mythos 5.1. Around it: a fullscreen diff panel, /skill-doctor, managed MCP servers pushed by admins, a subagent-model force switch, and a round of auto-mode hardening. Continues Week 35.

The official W36 digest (dated August 31 – September 4, releases v2.1.251 → v2.1.261) headlines four features: Claude Fable 5.1, background computer use in the Desktop app, the live /diff panel, and /skill-doctor. Its range starts at v2.1.251, which shipped on 2026-08-28, so several of its “other wins” — the PreModelSwitch/PostModelSwitch hooks, the /cost prompt-cache line, per-model /effort, the Enterprise → Opus 5 default — are v2.1.251 items recorded in Week 35.

Key Takeaways

  • Claude Fable 5.1 (claude-fable-5-1) is the default Fable model — 1M context, 50 per Mtok, $0.25/Mtok cache reads (v2.1.257, 2026-09-01). See Claude Fable 5.1 + Mythos 5.1.
  • Gateway users stay on Fable 5 for now. In Claude apps gateway sessions, fable and best keep resolving to Fable 5, because gateways not yet configured for Fable 5.1 reject it; pick Fable 5.1 in /model to use it (v2.1.257).
  • Fable 5.1 launch fixes landed two days later (v2.1.260): the /model picker now shows Fable 5.1 to eligible organizations (it was only accepted when typed as /model claude-fable-5-1); prompt caching now covers context attached after tool results (it was re-sent uncached on every tool-call turn); changing /effort mid-session no longer invalidates the cache.
  • The claude-api skill lists three breaking changes from Fable 5 to Fable 5.1: a forced tool_choice of any/tool returns 400; thinking blocks are preserved only for the model that produced them or a newer one; and only in the conversation that produced them (skills 53048666).
  • Diff panel beside the conversation in fullscreen mode shows uncommitted changes as Claude edits; toggle with /diff (v2.1.260). Select lines with the mouse to attach them to your next prompt.
  • Background computer use in the Desktop app (official digest; macOS, beta on Pro and Max): Claude sees and acts in the apps you’ve approved while you keep working.
  • /skill-doctor shows which loaded skills go unused and what they cost in context (v2.1.261). See Agent Skills.
  • managedMcpServers lets organizations push HTTP/SSE MCP servers to every user, and allowedMcpServers now governs only servers users add (v2.1.259). See MCP.
  • CLAUDE_CODE_SUBAGENT_MODEL_FORCE applies CLAUDE_CODE_SUBAGENT_MODEL (or the main model) to every subagent, ignoring per-spawn and agent-definition overrides — the escape hatch for W35’s change of that variable into a default (v2.1.257).
  • defaultMode: "bypassPermissions" in project settings is now ignored, like "auto" already was; set it in user or managed settings or pass --permission-mode (v2.1.257).
  • Auto mode gets a Containment Escape rule — cloud metadata-credential fetches, egress evasion and cross-tenant reach are no longer auto-approved unless the environment marks them expected — plus a one-time prompt before the first read outside the working directories (v2.1.257). The digest’s examples: requesting credentials from the cloud instance-metadata endpoint, and connecting to sibling containers Claude didn’t start. See Auto Mode.

Release table

VersionDate (UTC)Shape
v2.1.2522026-08-31Fixes only
v2.1.253–256—No GitHub release pages (see Open Questions)
v2.1.2572026-09-01Claude Fable 5.1; Containment Escape rule; timeFormat/timeZone; subagent model force
v2.1.2582026-09-01Fixes only (macOS 12 launch regression from 2.1.255)
v2.1.2592026-09-02managedMcpServers, --permission-prompts none, GitLab MR recognition
v2.1.2602026-09-03Diff panel, Fable 5.1 fixes, cache-miss causes, text /advisor
v2.1.2612026-09-04/skill-doctor, bashOutputMaxChars, Bash-style word keys
v2.1.262—No GitHub release page
v2.1.2632026-09-06”Bug fixes and reliability improvements” only

Models and defaults

  • Claude Fable 5.1 added and made the default Fable model (v2.1.257).
  • Gateways: fable/best → Fable 5 “for now” (v2.1.257).
  • Fable 5.1 follow-ups (v2.1.260): /model picker visibility; caching after tool results; model: fable agents now honour [1m] on an ANTHROPIC_DEFAULT_FABLE_MODEL pin (they were silently running with a 200K window); /effort changes no longer break the cache.
  • Auto-compact for 1M-context models: Opus and Fable sessions now compact shortly before the 1M-token limit, and recovery compaction on very large contexts no longer times out at 10 minutes (v2.1.260).
  • s in /effort changes effort for the current session only, matching /model (v2.1.257); --effort lifts a new model’s default-effort hold for that session only rather than permanently (v2.1.257).
  • Frontmatter model: on custom commands and skills was ignored in interactive sessions — fixed (v2.1.259); auto mode no longer runs a turn on an unsupported model named in command or skill frontmatter (v2.1.259).
  • The bundled claude-api skill’s Go, Java and C# samples use current-generation model IDs, and the skill clarifies that worker and sub-agent models should also be current-generation (v2.1.260).
  • The /model picker shows friendly names instead of raw Bedrock, Vertex or gateway IDs when it recognizes them (v2.1.261).
  • Bedrock and Bedrock Mantle streams on Opus 4.7 and later carry progress events during long hidden-thinking phases, so idle timeouts no longer cut the connection (v2.1.257).

The claude-api skill update (skills 53048666, 2026-09-01)

Four squashed commits in anthropics/skills updated the bundled claude-api skill for the Fable 5.1 launch:

  • Claude Fable 5.1 (claude-fable-5-1) and Claude Mythos 5.1 (claude-mythos-5-1) become the default Fable-tier models throughout the skill; Fable 5 and Mythos 5 stay selectable by ID with their own catalog rows.
  • A “Migrating to Claude Fable 5.1 from Claude Fable 5” section with the three breaking changes above, what carries over, the Opus 5 path, Mythos 5.1 notes, and a migration checklist.
  • New API features documented: per-message effort (mid-conversation-output-config beta); turn-scoped mid-conversation system messages with clear_at; progress updates between tool calls via thinking.display "updates"; thinking block_binding controls; and a 0.025x cache-read rate on Fable 5.1 with a max_tokens: 0 keep-alive the commit says “usually beats the 1-hour TTL”.
  • Files API and Skills API are out of beta: examples use client.files.* / client.skills.* with no beta header.
  • Sonnet 5’s 10 is the permanent list price; advisor pairing no longer excludes Claude Fable 5 for Managed Agents; new cost-optimize and Admin API references.
  • Launch-day hedges the commit leaves open: Task Budgets support on Fable 5.1, whether Mythos 5.1 shares the 0.025x cache-read rate, and the fallback-credit wording.

A separate commit (41bbe19d, 2026-09-03) updated the frontend-design skill “to avoid generic design defaults”; the commit body gives no further detail.

From the official W36 digest

The digest carries detail the release notes don’t, and one headline feature (background computer use) that ships in the Desktop app rather than in a CLI release.

Background computer use in the Desktop app (macOS, beta)

  • On macOS, computer use in the Claude Code Desktop app now works in the background: Claude sees and acts in the apps you’ve approved while you keep working. Beta on Pro and Max plans; Desktop only. Reference: code.claude.com/docs/en/desktop#let-claude-use-your-computer.
  • The practical change is that a computer-use task no longer needs the screen to itself, so you can keep working in other apps while it runs.

Live /diff panel (v2.1.260)

  • In fullscreen rendering, /diff opens a panel beside the conversation instead of a viewer you have to close. It lists changed files with added and removed line counts and refreshes each time Claude edits a file or runs a shell command.
  • Select lines in the panel with the mouse to attach them to your next prompt.
  • Requirements: fullscreen rendering on, inside a git repository, terminal at least 110 columns wide. Close it with /diff again or the X in its header.

/skill-doctor

  • Every skill in the skill listing adds to context on every turn, whether or not Claude uses it; /skill-doctor shows each skill’s context cost and how often it gets used.
  • In an interactive session it opens the report in the /plugin manager’s Stats tab; with -p it prints the report as text. It isn’t available in sessions that skip feature-flag fetching.
  • The digest says it “requires v2.1.252 or later”; the GitHub release notes list it as added in v2.1.261.

Fable 5.1 in the /model picker

  • /model fable switches the session to Fable 5.1 and saves it as your default. On the Anthropic API, the picker lists Fable only once the server reports it available for your organization, but typing /model fable checks with the server directly. Compare the v2.1.260 fix in Key Takeaways, after which the picker shows Fable 5.1 to eligible organizations.

Other wins

  • VS Code: click the model name at the bottom of the prompt box to open the model picker — the footer model pill added in v2.1.257, with an Effort row and a “More models” page; select Output styles in the command menu’s Customize section to pick an output style, including custom ones (v2.1.257).
  • The remaining items (Containment Escape examples, the out-of-directory read prompt, managedMcpServers, bashOutputMaxChars, readline word keys, the bypassPermissions restriction, the s key in /effort) are folded into the sections below; the v2.1.251 items are in Week 35.

New commands and features

  • Diff panel in fullscreen mode, toggled with /diff (v2.1.260).
  • /skill-doctor (v2.1.261).
  • /cost and the status line’s prompt_cache field report a likely cause for cache misses, e.g. tool definitions or system prompt changed, or idle past the TTL (v2.1.260). The digest calls the line Prompt cache (main): the share of input tokens served from cache, the misses, whether the cache is warm, and a likely cause for the last miss when Claude Code can name one.
  • Text form of /advisor (/advisor, /advisor <model>, /advisor off) for the desktop app, Remote Control and headless -p/Agent SDK sessions (v2.1.260). See Advisor Tool.
  • /reload-plugins in headless sessions, so it appears in Claude Code Desktop and SDK command lists (v2.1.260).
  • --permission-prompts none for unattended headless hosts: anything that would prompt is denied automatically while the active permission mode, including auto mode, keeps deciding (v2.1.259).
  • --append-subagent-system-prompt-file for subagent system prompts too large for the command line (v2.1.261).
  • claude plugin validate --json (v2.1.259).
  • GitLab: glab mr create/merge/close/reopen/note/update are recognized and show as MR !N in the tool summary and footer badge (v2.1.259); /code-review --comment posts findings to GitLab merge requests via glab mr note (v2.1.257).
  • Auto mode: Containment Escape rule and the out-of-working-directory read prompt with permissions.blockReadsOutsideWorkingDirectories (v2.1.257); a link that packs content into a public diagram renderer’s URL is treated as an upload and not auto-approved (v2.1.261).
  • /fork keeps the original conversation’s prompt cache in the new background session (v2.1.257).
  • Background commands started by subagents no longer have a one-hour limit (v2.1.260).
  • /ultrareview waits up to 45 minutes for long cloud reviews, up from 30 (v2.1.260). See Ultrareview.

Settings and environment variables

  • timeFormat (“Time format”) and timeZone: 12-hour, 24-hour, 24-hour UTC or a strftime pattern for the turn-end clock and transcript timestamps (v2.1.257).
  • managedMcpServers (v2.1.259); allowedMcpServers narrowed to user-added servers — a managed-mcp.json server your allowlist used to filter out now loads on upgrade; use deniedMcpServers to keep it off (v2.1.259).
  • bashOutputMaxChars and taskOutputMaxChars raise inline command and background-task output up to 128K characters before it spills to a file (v2.1.261). (taskOutputMaxChars stopped having any effect in v2.1.277 when the TaskOutput tool was removed — see Week 38.)
  • defaultMode: "bypassPermissions" ignored in project .claude/settings.json and .claude/settings.local.json; the session starts in Manual mode instead (v2.1.257).
  • managedSourcesBehavior: "merge" takes sandbox.credentials.awsPairs and sandbox.ripgrep whole from the highest managed source that sets them (v2.1.257).
  • A managed CLAUDE.md (claudeMd) no longer triggers the security approval dialog; hooks, shell-command, sandbox and unsafe env settings still do (v2.1.260).
  • Commands typed at the ! bash-mode prompt run outside the sandbox even under strict sandbox mode (v2.1.260).
  • Permission rules with text after the closing parenthesis (e.g. Bash(ls) x) are reported as invalid instead of silently ignored (v2.1.260).
  • keybindingFlavor no longer has any effect: the prompt’s word-editing keys now match Bash for everyone (Ctrl+W, Alt+F, Alt+D, punctuation separates words) (v2.1.261). The digest’s wording: Ctrl+W deletes back to the previous whitespace, and Alt+B, Alt+F and Alt+D treat punctuation such as / and . as word breaks. The setting was added in Week 34 (v2.1.238).
  • Managed settings that can’t be parsed make Claude Code refuse to start, naming the source (v2.1.259).
  • Machines whose managed settings pin forceLoginMethod: "gateway" ignore a leftover API key or claude.ai login and ask for /login (v2.1.261).
  • Claude apps gateway: oidc.scope_on_refresh; desktop policy keys userPluginMarketplacesEnabled, userPluginUploadsEnabled (v2.1.260).

Hooks, MCP, plugins, skills and subagents

  • Blocking Stop hooks no longer cause the next turn to lose the model’s reasoning or, on some models, miss the prompt cache (v2.1.259). See Hooks.
  • Resumed sessions no longer lose hook output around parallel tool calls (v2.1.261).
  • Plugins can no longer read outside their own directory through a symlinked component path (v2.1.257).
  • Managed skillOverrides keyed on a bundled skill’s alias now apply, and Skill(name) deny rules cover nested <dir>:name skills (v2.1.260).
  • Model switching no longer stays blocked for the session after a plugin hook load failure (v2.1.260).
  • MCP: servers that disconnect while listing tools now report the error (v2.1.259); SDK-provided MCP servers are present on the first turn (v2.1.260); claude.ai connectors are retried in the background after a startup timeout (v2.1.261).
  • Workflow agent({schema}) rejects an unsatisfiable JSON Schema up front (v2.1.260); /workflows pretty-prints JSON outcomes (v2.1.259).
  • Subagents automatically continue after a mid-stream cut (v2.1.257); a subagent that resumed another via SendMessage is now woken on its completion (v2.1.260); nested background subagent results are saved in the parent subagent’s transcript (v2.1.259). See Subagents.
  • Agent teams: in-process teammates no longer re-send first-turn tool and skill announcements, which had broken the prompt cache (v2.1.261).

Breaking and behaviour changes

  • /btw history browsing moved to Shift+←/Shift+→ (or [/]) (v2.1.257).
  • ctrl+l / cmd+k in fullscreen clear the transcript view (v2.1.260) — reverted in v2.1.280 (Week 39).
  • The Ctrl+E command explanation on Bash and PowerShell permission prompts is removed (v2.1.257).
  • --add-dir and additionalDirectories refuse network paths (UNC shares, /net/<host> automounts) (v2.1.257).
  • Cowork and claude.ai cloud sessions: reading an artifact that isn’t yours always asks first, even in auto mode (v2.1.257).
  • Claude in Chrome follows the organization’s admin setting; when off, --chrome, /chrome and the browser tools are unavailable (v2.1.260).
  • claude --resume <id> --bg continues that session under its own ID when nothing is running it (v2.1.257).
  • v2.1.259 applied Read() deny rules to Bash arguments (option values, git diff/git grep operands, cd DIR && cat FILE); v2.1.260 reverted it because it denied npm run build under a Read(./**/build/**) rule and made cd … && grep prompt even in auto mode.

Notable fixes

  • Concurrent sessions silently reverting each other’s ~/.claude.json changes — trust resets and lost MCP state (v2.1.259).
  • A permissions.ask rule skipped in auto mode inside compound commands or subshells (v2.1.257).
  • File permission rules containing parentheses dropped or ignored by the Bash sandbox, leaving “read-only” folders writable (v2.1.260).
  • Prompt cache: advisor-model sessions missing the cache on background requests (v2.1.257); screenshot-heavy sessions missing it every turn (v2.1.257).
  • v2.1.258: Claude Code failing to launch on macOS 12 Monterey, “a regression introduced in 2.1.255”.
  • v2.1.252: “task output swap refused” Bash failures on some Macs; “always allow” not saving in a project without .claude/settings.local.json.

Try It

  • Gateway users: pick Fable 5.1 explicitly in /model if your gateway is configured for it; fable and best still mean Fable 5 there.
  • Update past v2.1.260 before judging Fable 5.1 cost in Claude Code. Before that release, context attached after tool results was re-sent uncached on every tool-call turn.
  • Run /skill-doctor and prune skills that cost context without being used.
  • Switch to Fable 5.1 with /model fable on the Anthropic API; if the picker doesn’t list it yet, typing the command checks availability with the server.
  • Try /diff in fullscreen in a git repo, then select lines in the panel to scope your next prompt to those changes.
  • If you relied on bypassPermissions in a project settings file, move it to user or managed settings.
  • For unattended CI hosts, try --permission-prompts none with auto mode instead of --dangerously-skip-permissions.

Open Questions

  • v2.1.253–256 and v2.1.262 have no GitHub release pages, yet v2.1.258 fixes “a regression introduced in 2.1.255”, so 2.1.255 shipped somewhere. The distribution channel is not stated.
  • v2.1.263 is unitemized (“Bug fixes and reliability improvements”).
  • When will gateways resolve fable/best to Fable 5.1? v2.1.257 says “for now” with no date.
  • What does the frontend-design skill change actually do? Commit 41bbe19d says only “avoid generic design defaults”.
  • Three claude-api skill claims were hedged at launch (Task Budgets on Fable 5.1, Mythos 5.1’s cache-read rate, fallback-credit wording); none of this week’s sources confirms them.
  • Which macOS versions support background computer use? The digest says only “macOS”; no minimum version or Apple Silicon/Intel split is given.
  • Does /skill-doctor report per skill or per plugin for a plugin that ships several skills?
  • Is bashOutputMaxChars applied per command or across a session’s Bash output? The notes describe output from “a successful command” without saying.
  • Which release shipped /skill-doctor — v2.1.252 (the digest’s minimum version) or v2.1.261 (the release note)?