Archive — 2026-04Week 4

Elevated errors on Claude Sonnet 4.5 — status.claude.com began investigating at 04-28 13:29 UTC and resolved at 13:50 UTC; actual impact window 13:22–13:39 UTC, lasting only about 17 minutes.

Elevated errors on Claude Haiku 4.5 — status.claude.com began investigating at 04-28 12:24 UTC, entered monitoring at 12:38 UTC, and resolved at 12:45 UTC; actual impact window 11:53–12:44 UTC, lasting approximately 51 minutes.

Anthropic launches Claude for Creative Work — debuts MCP-based connectors for creative tools: Ableton (grounds responses in official Live and Push docs), Adobe Creative Cloud (50+ tools including Photoshop and Premiere), Autodesk Fusion (conversational 3D modeling), Blender (natural-language interface to its Python API, scene analysis and custom script generation), Affinity by Canva (batch processing automation), Splice (royalty-free sample search), Resolume and SketchUp (real-time control and 3D modeling); connectors are MCP-based so they are also accessible from other LLMs; Anthropic also joined the Blender Development Fund as a patron and partnered with RISD, Ringling College, and Goldsmiths to integrate the tools into creative curricula.

Claude Code 2.1.121 -- Adds `alwaysLoad` option to MCP server config (when `true`, all tools from that server skip tool-search deferral and are always available); adds `claude plugin prune` to remove orphaned auto-installed plugin dependencies (`plugin uninstall --prune` cascades); `/skills` gains a type-to-filter search box; PostToolUse hooks can now replace tool output for all tools via `hookSpecificOutput.updatedToolOutput` (previously MCP-only); `--dangerously-skip-permissions` no longer prompts for writes to `.claude/skills/`, `.claude/agents/`, and `.claude/commands/`; MCP servers that hit transient startup errors now auto-retry up to 3 times; `/terminal-setup` enables iTerm2's clipboard-access setting so `/copy` works from tmux; fixes unbounded memory growth (multi-GB RSS) when processing many images, `/usage` leaking up to ~2GB on large transcript histories, Bash tool becoming permanently unusable when its working directory is deleted/moved mid-session, `--resume` crashing in external builds, `--resume` failing on large sessions with corrupted transcript lines (now skipped), and scrollback duplication on Ctrl+L or redraw in tmux/GNOME Terminal/Windows Terminal/Konsole; fullscreen mode no longer jumps to the bottom when you type, dialogs overflowing the terminal are now scrollable with arrow keys / PgUp/PgDn / mouse wheel; [VSCode] `/context` opens a native token usage dialog, voice dictation respects `accessibility.voice.speechLanguage`.

Elevated errors on Claude Opus 4.7 -- status.claude.com began investigating at 04-25 08:43 UTC, entered monitoring at 08:57 UTC, and resolved at 11:58 UTC; duration approximately 3 hours 15 minutes; affected claude.ai, platform.claude.com, API, and Claude Code.

@ClaudeDevs announces a significant refresh of Claude Code on the web and the iOS app: (1) new UI, (2) improved speed and reliability, (3) cross-device workflow -- start a session on web or mobile, then run `claude --teleport` in your terminal to continue with the full conversation and branch state. Live now at claude.ai/code and on iOS.

Anthropic blog "Anthropic and NEC partner to build AI-native engineering organization" — NEC becomes Anthropic's first Japan-based global partner. Approximately 30,000 NEC Group employees worldwide will deploy Claude as NEC builds one of Japan's largest AI-native engineering organizations. The companies will jointly develop secure, domain-specific AI products for the Japanese market, with an initial focus on finance, manufacturing, local government, and cybersecurity. NEC is establishing a Center of Excellence to grow its AI-native engineering team using Claude Code and Claude Cowork, taking a "Client Zero" approach where NEC is its own first customer. On the security side, NEC is integrating Claude into NEC BluStellar Scenario and its next-generation cybersecurity SOC services. The NEC official press release was issued on 2026-04-23.

Google commits up to $40 billion in Anthropic -- an initial $10 billion cash injection, with up to $30 billion more contingent on hitting performance milestones. Google Cloud simultaneously commits 5 gigawatts of computing capacity over five years, with an option for several additional gigawatts. Anthropic's annualized run-rate revenue has surpassed $30 billion.

Anthropic releases the Rate Limits API, allowing administrators to programmatically query the rate limits configured for their organization and workspaces.

Claude Code 2.1.119 -- `/config` settings (theme, editor mode, verbose, etc.) now persist to `~/.claude/settings.json` with project/local/policy override precedence; adds `prUrlTemplate` setting (custom code-review URL for footer PR badge) and `CLAUDE_CODE_HIDE_CWD` env var (hides working directory in startup logo); `--from-pr` now accepts GitLab merge-request, Bitbucket pull-request, and GitHub Enterprise PR URLs; `--print` mode honors agent `tools:`/`disallowedTools:` frontmatter; `--agent` honors the agent definition's `permissionMode`; PowerShell tool commands can be auto-approved in permission mode; `PostToolUse`/`PostToolUseFailure` hook inputs now include `duration_ms`; plugin version constraints auto-update to the highest satisfying git tag; Vim INSERT: Esc no longer pulls a queued message back into the input; slash command suggestions highlight matched characters and wrap long descriptions. Fixes: CRLF paste (Windows clipboards, Xcode console) inserting extra blank lines between lines; fullscreen scroll snapping to bottom after a tool finishes; MCP HTTP OAuth discovery failing on non-JSON body; `blockedMarketplaces` `hostPattern`/`pathPattern` not correctly enforced; `${ENV_VAR}` placeholders in MCP server headers not substituted; MCP OAuth client secret not sent during `client_secret_post` token exchange; `/skills` Enter closing the dialog instead of pre-filling the skill name; `/agents` detail view mislabeling built-in tools as 'Unrecognized'; stray `p` character at prompt in macOS Terminal.app via Docker/SSH; `TaskList` returning tasks in arbitrary order instead of sorted by ID; PR not linked to session when working in a git worktree; and many more.

Anthropic publishes an official post-mortem on the recent Claude Code quality regression -- confirming users were not imagining the drop in intelligence over the past month. Three product-layer changes were responsible: (1) On Mar 4, default reasoning effort was lowered from `high` to `medium` to reduce UI latency (reverted Apr 7); (2) On Mar 26, a caching optimization that cleared older thinking from sessions idle for over an hour had a bug causing it to trigger every turn instead of just once (fixed Apr 10); (3) On Apr 16, system prompt instructions were added to limit verbosity (under 25 words between tool calls, under 100 words for final responses), which hurt Opus 4.7 coding evaluations (reverted Apr 20). Affected Sonnet 4.6 / Opus 4.6 / Opus 4.7. Anthropic also reset all subscribers' 5-hour and weekly usage quotas as compensation.

Memory for Claude Managed Agents is now in public beta. Memory stores allow agents to persist information across sessions; each store is a workspace-scoped collection of text documents mounted as a directory under `/mnt/memory/` inside the session container, readable and writable using the standard agent toolset. Every mutation creates an immutable memory version for audit and rollback (retained 30 days). A maximum of 8 memory stores can be attached per session with `read_write` or `read_only` access controls. Requires the `managed-agents-2026-04-01` beta header.

Claude Code 2.1.118 -- Adds vim visual mode (`v` for character selection, `V` for line selection, with operators and visual feedback); merges `/cost` and `/stats` into `/usage` (the original commands remain as typing shortcuts that open the relevant tab); create or switch named custom themes from `/theme`, or hand-edit JSON in `~/.claude/themes/`; plugins can ship themes via a `themes/` directory; Hooks gain `type: "mcp_tool"` to invoke MCP tools directly; adds `DISABLE_UPDATES` env var (stricter than `DISABLE_AUTOUPDATER`, also blocks manual `claude update`); Auto mode lists support `"$defaults"` token to append custom rules on top of the built-in list; fixes multiple MCP OAuth issues (token responses missing `expires_in` forcing re-auth every hour, step-up authorization silently refreshing instead of prompting, cross-process lock contention, macOS keychain race overwriting freshly-refreshed tokens with misleading `/login` prompts, refresh failing when the server revokes a token early); fixes Linux/Windows credential save crash corrupting `~/.claude/.credentials.json`, `/fork` writing the full parent conversation per fork, Alt+K / Alt+X freezing keyboard input, and many more.

Claude.ai MCP applications unavailable -- status.claude.com identified the issue at 04-23 00:41 UTC and resolved it at 02:09 UTC; duration approximately 1 hour 28 minutes; affected claude.ai, platform.claude.com, API, Claude Code, Cowork, and Claude for Government.

Anthropic briefly A/B tested removing Claude Code from the Pro plan for 2% of new Pro subscribers -- the pricing page temporarily showed Pro ($20/month) without Claude Code, but following strong community backlash, it was restored within hours. Anthropic acknowledged a surge in usage and said the plan structure is under review.

Claude Code 2.1.117 -- Adds `CLAUDE_CODE_FORK_SUBAGENT=1` to enable forked subagents on external builds; `/model` selections persist across restarts (even when the project pins a different model); `/resume` now offers to summarize stale, large sessions before re-reading; local and claude.ai MCP servers connect concurrently for faster startup; macOS/Linux native builds replace `Glob`/`Grep` with embedded `bfs`/`ugrep` invoked through the Bash tool (Windows and npm installs unchanged); default effort for Pro/Max subscribers on Opus 4.6 / Sonnet 4.6 raised from `medium` to `high`; fixes OAuth tokens expiring mid-session and forcing manual `/login`, WebFetch hanging on very large HTML pages, and Opus 4.7 `/context` percentages computed against 200K instead of the actual 1M window.