Archive — 2026-05Week 5

Claude Opus 4.7 elevated error rate — status.claude.com began investigating at 22:58 UTC on 2026-05-30 and marked it resolved at 00:16 UTC on 2026-05-31, lasting about 1 hour 18 minutes; impact was limited to claude.ai and Claude Console (platform.claude.com); Anthropic did not disclose a root cause. This is the fifth Opus 4.7 error-rate incident since 5/25 (5/25, two on 5/27, 5/28, 5/30), occurring after Opus 4.8 went live on 5/28.

Claude Code 2.1.158 — Auto mode is now available on Bedrock, Vertex, and Foundry for Opus 4.7 and Opus 4.8; opt in by setting `CLAUDE_CODE_ENABLE_AUTO_MODE=1`.

Claude Code 2.1.157 — New: plugins in `.claude/skills` directories are auto-loaded (no marketplace required); added `claude plugin init <name>` to scaffold a plugin in `.claude/skills`; autocomplete for `/plugin` arguments (subcommands, installed plugin names, plugins from known marketplaces); `claude agents` honors the `agent` field in `settings.json` for dispatched sessions, with `--agent <name>` to override; `EnterWorktree` can switch between Claude-managed worktrees mid-session; `tool_decision` telemetry events include `tool_parameters` (bash commands, MCP/skill names) when `OTEL_LOG_TOOL_DETAILS=1`; Claude-managed worktrees are left unlocked after the agent finishes so `git worktree remove`/`prune` can clean them up. Fixes: unprocessable images (zero-byte, corrupt) attached via paste/MCP/dialog become a text placeholder instead of crashing the request; sandbox network permission prompts appearing in auto and bypass-permissions mode on desktop app/IDE extensions/SDK; `claude agents` completed sessions not retiring when an idle subagent was parked or had leaked a backgrounded shell; Esc not cancelling a slow "opening…" leaving the list unresponsive; background agent worktrees orphaned after the 30-day job retention sweep; background sessions re-attached after sleep/wake not telling the model the correct date; copy-on-select in `claude agents` not reaching the system clipboard under tmux `set-clipboard on` (2.1.153 regression); `--resume` not reporting background subagents running when the previous process exited; `--worktree`/`--worktree --tmux` returning to the canonical repo root instead of the current linked worktree; the `/model` picker showing a wrong "Newer version available" hint when the selected model is already newest in its family, with the pinned row now showing the description instead of the raw ID; literal markdown markers in in-progress message text in fullscreen mode; terminal freezing after approving the managed-settings security dialog at startup; a rare duplicate line in scrollback after UI redraw; right-click paste duplicating the clipboard in VS Code/Cursor/Windsurf terminals; WSL image paste (`alt+v`), Windows 11 screenshot paste, and dragging images from Windows Explorer; IDE Stop not actually stopping a running background subagent; the fast mode indicator not appearing on Opus 4.8 in VSCode. Improvements: faster long and resumed conversations (eliminated redundant message-render recomputation); `/terminal-setup` disables GPU acceleration in integrated terminals to prevent garbled text; the Feature of the Week credit-claim status now shows as a status-area notification; removed the "bash commands will be sandboxed" and "/ide for …" startup hints; backspace right after a workflow trigger keyword now dismisses the workflow request (same as alt+w); a new "Workflow keyword trigger" setting in `/config` can stop the word "workflow" from triggering a dynamic workflow.

Anthropic announced via the official @ClaudeDevs X account a new Claude API behavior from Opus 4.8 onward — developers can now pass a `system`-role message mid-conversation, and Claude treats it as an authoritative instruction from that point forward. Because prior system instructions remain in the prompt cache, adding a mid-conversation system message does not break the cache and preserves cache hits, lowering cost and latency for API requests. This behavior is supported only from Opus 4.8 onward, with documentation updated in the automatic-caching section of prompt-caching at platform.claude.com.

Elevated errors on Claude Opus 4.8 (second incident the same day, 5/29) — status.claude.com began investigating at 18:35 UTC on 2026-05-29, deployed a fix with success rates back to normal and entered monitoring at 18:55 UTC, and marked it resolved at 19:12 UTC; the error impact window was roughly 18:16–18:42 UTC (about 26 minutes); scope was the Claude Opus 4.8 model; Anthropic did not disclose a root cause. This was the second Opus 4.8 error-rate incident on 5/29 (the first started 08:30 UTC and resolved 08:45 UTC, about 15 minutes) and the second since the model launched on 5/28.

Elevated errors on Claude Opus 4.8 — status.claude.com began investigating at 08:30 UTC on 2026-05-29, applied a fix and moved to monitoring at 08:39 UTC, and resolved the incident at 08:45 UTC, lasting roughly 15 minutes; the impact spanned claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic has not disclosed a root cause. This is the first error-rate incident for Opus 4.8 (launched 5/28), following the string of similar Opus 4.7 alerts over the preceding days.

Salesforce's official blog published "Pioneering the Agentic Shift Within Salesforce Engineering" — describing the engineering organization's pivot from "AI as copilot" to "agentic tools driving the SDLC": Claude Code was rolled out to all engineers with token limits removed; built the AI Expert Suite and Salesforce Foundation Plugins as a curated, institutionalized skill library; subagents handle parallel workstreams within larger tasks so engineers no longer context-switch across five systems for a single task. Concrete metrics: a migration scoped at 231 days shipped in 13; one PR delivered 21 endpoints at 100% test coverage; total incidents dropped 5% even as PR throughput rose substantially (security guardrails and quality standards built directly into the agentic workflow). Boris Cherny, creator and head of Claude Code, amplified on X: "the teams seeing the biggest wins from AI are completely changing how they work, not speeding up what they already do — what steps can you delete, what handoffs go away, what can an agent just own end-to-end." The post also looks ahead at challenges: context management, security considerations, and role evolution as agents absorb more of the execution layer.

Claude Code 2.1.156 — Fixed an issue where, when using Opus 4.8, thinking blocks were modified, leading to API errors.

Claude billing and subscription management errors — status.claude.com marked the issue identified at 19:04 UTC, deployed a fix and began monitoring at 19:17 UTC, and marked it resolved at 19:23 UTC on 2026-05-28, lasting about 19 minutes. Impact was limited to billing and subscription management on claude.ai and the Anthropic Console (errors when viewing subscription details, managing payment methods, or upgrading/changing plans); chat and API access were unaffected. Anthropic did not disclose a root cause.

Claude Opus 4.8 is generally available on GitHub Copilot — GitHub's official changelog announced Anthropic's latest flagship, Claude Opus 4.8, marking a clear step forward in code understanding and generation, complex problem-solving, and large-codebase navigation, now generally available to Copilot Pro+, Business, and Enterprise plans. Supported editors span Visual Studio Code (chat, ask, edit, and agent modes), Visual Studio, Copilot CLI, the GitHub Copilot cloud agent, the GitHub Copilot App, github.com, GitHub Mobile (iOS and Android), JetBrains, Xcode, and Eclipse. It launches with a 15X premium request multiplier until usage-based billing begins on 2026-06-01; Enterprise and Business administrators must enable the Claude Opus 4.8 policy in Copilot settings to grant access.

Claude Code 2.1.154 — Added Claude Opus 4.8 support, now defaulting to high effort with `/effort xhigh` for the hardest tasks; introduced Dynamic Workflows — ask Claude to create a workflow and it orchestrates tens to hundreds of agents in the background so you can tackle larger, more complex tasks (`/workflows` to view runs); Opus 4.8 fast mode is now 2x the standard rate for 2.5x the speed; the lean system prompt is now the default for all models except Haiku, Sonnet, and Opus 4.7 and earlier; Claude now reserves the multiple-choice prompt for decisions it genuinely cannot make itself, proceeding directly when it has enough context; `/simplify` now runs a cleanup-only review (reuse, simplification, efficiency, altitude) and applies the fixes instead of the full `/code-review --fix` bug hunt; the `/effort` slider labels were renamed from "Speed"/"Intelligence" to "Faster"/"Smarter"; `claude agents` can run a shell command as an attach/detachable background session via `! <command>` (or `claude --bg --exec`); plugins can declare `defaultEnabled: false`, and the Discover tab pins directory-relevant plugins as "suggested for this directory"; streaming tool execution is now always enabled (including with telemetry disabled and on Bedrock/Vertex/Foundry); the auto-mode classifier has improved data-exfiltration detection, especially bulk transfers of repository contents. Security fixes: `rm -rf $HOME` not blocked as a dangerous path when HOME has a trailing slash, `$TMPDIR` resolving differently in sandboxed vs unsandboxed Bash, subagents in background sessions bypassing worktree isolation to write the shared checkout, orphaned `claude --bg-pty-host` spinning at 100% CPU on macOS after the daemon exits, and pinned background sessions respawning every minute after an update causing agent-start notification floods. Deprecated `CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE` (removed 06/01; use `/model claude-opus-4-6[1m]` + `/fast on` instead); added Opus 4.8 support and 4.7→4.8 migration guidance to the `/claude-api` skill.

Reuters reported that Anthropic launched an upgraded Claude Opus 4.8 on Thursday (5/28) while saying it is working to release its "market-moving" Mythos model to all customers in the coming weeks. Opus 4.8 is priced the same as its predecessor and shows improvements across benchmarks, particularly in honesty: early testers report it is more likely to flag uncertainties around its work and less likely to make unsupported claims. Mythos is the company's large language model with advanced cybersecurity capabilities whose impact has raised concerns among executives and world leaders; as part of Project Glasswing, major tech firms including Amazon, Microsoft, and Apple are permitted to use Mythos for cybersecurity purposes. Context: because of its nation-state-grade cyberattack capabilities, Anthropic had chosen not to release Mythos publicly, granting access only to a small set of partners and governments for testing.

Sid Bidasaria (@sidbid), the lead engineer behind Anthropic's Dynamic Workflows, posted a builder's tips thread on X (eight posts) complementing the @ClaudeDevs 5/28 launch thread with trigger and usage details not covered there: (1) **Trigger mechanism** — mentioning the word "workflow" in any prompt causes Claude to spin one up automatically; `/effort ultracode` mode makes Claude use workflows aggressively without being asked. (2) **When to use** — tasks too large for a single session, such as codebase-wide bug hunts, and @jarredsumner's actual rewrite of Bun from Zig to Rust. (3) **Adversarial judges** — route work through multiple judges that each attack from a different lens for high-confidence critical tasks; can scale to a tournament-style bracket (e.g., a plan-hunter workflow scoring plans). (4) **Reusable** — workflows are orchestration scripts under the hood — save them, commit them to your repo, share them with your team, or turn them into skills. (5) **Cost warning** — parallel agents make workflows expensive — try small examples first to gauge token usage; can be disabled via `/config` or enterprise settings. Sid notes "no other coding tool I know of lets you do this," and that the feature has already been a daily driver for a subset of Anthropic engineers, built over the past "couple of months."

Anthropic introduced Claude Opus 4.8 — its most capable generally available model, succeeding Opus 4.7 with across-the-board gains in coding, agentic, and knowledge work, and a particular emphasis on honesty: early testers report it is more likely to flag uncertainties about its work and less likely to make unsupported claims, addressing the common failure where models jump to conclusions and claim progress on thin evidence. Official benchmarks vs Opus 4.7: agentic coding 64.3%→69.2%, knowledge work 1753→1890, agentic financial analysis 51.5%→53.9%, agentic computer use 82.8%→83.4%. Pricing is unchanged from Opus 4.7 — $5 per million input tokens and $25 per million output tokens standard; fast mode is 2x the standard rate ($10/$50) for 2.5x the speed, roughly 3x cheaper than the previous generation's fast mode. Three features shipped alongside: Dynamic Workflows (research preview — orchestrate tens to hundreds of parallel subagents in a single background session for large tasks), Effort Control (selectable effort levels in claude.ai and Cowork to trade off token usage vs speed), and a Messages API update accepting system entries within the messages array for mid-task instruction updates. The model ID is `claude-opus-4-8`, available natively on the Claude Platform and on AWS, Google Cloud, and Microsoft Foundry. Anthropic also said it expects to bring Mythos-class models to all customers in the coming weeks, once stronger cyber safeguards are in place.