Archive — 2026-06Week 1

Elevated errors on Claude Opus 4.7 — status.claude.com began investigating elevated request error rates on Opus 4.7 at 14:35 UTC on 2026-06-07, marked the issue identified with a fix in progress at 14:42 UTC, implemented a fix and entered monitoring at 14:57 UTC, and resolved at 15:41 UTC, lasting about 1 hour 6 minutes; impact spanned claude.ai, Claude Console (platform.claude.com), the Claude API (api.anthropic.com), and Claude Code; Anthropic did not disclose a root cause. It is the second platform incident of the day after the 03:31 UTC multi-model degradation, continuing the pattern of recurring frontier-model error-rate incidents since the Opus 4.8 launch.

Degraded performance for multiple models — status.claude.com began investigating elevated request error rates on Opus 4.7 at 03:31 UTC on 2026-06-07, identified the root cause of elevated error rates across several Claude models with a fix in progress at 03:41 UTC, and resolved at 04:28 UTC (having fixed an issue causing elevated error rates on requests to some Claude models), lasting about 57 minutes; impact spanned claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause. The incident widened from a single Opus 4.7 investigation to multiple models, reflecting continued error-rate incidents since the Opus 4.8 launch.

Claude Code 2.1.168 — official release notes list only "bug fixes and reliability improvements" with no itemized details. Released again the same day after 2.1.167 (01:33 UTC) and 2.1.166 (00:55 UTC).

Elevated errors on Claude Opus 4.8 — status.claude.com flagged an Opus 4.8 elevated-error incident (affected window 18:12–18:55 UTC), stated the issue was identified and being resolved at 18:37 UTC, and marked it resolved at 19:05 UTC, lasting about 43 minutes; impact spanned claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause. It is the second Opus 4.8 error-rate incident of the day after the 09:27 UTC Opus 4.8 degradation.

Opus 4.8 degraded service — status.claude.com began investigating an Opus 4.8 service degradation at 09:27 UTC on 2026-06-06, marked the issue identified with a fix in progress at 09:43 UTC, implemented a fix and entered monitoring at 09:59 UTC, and resolved at 10:14 UTC, lasting about 47 minutes; impact spanned claude.ai, Claude Console (platform.claude.com), the Claude API (api.anthropic.com), and Claude Code; Anthropic did not disclose a root cause. It is another Opus 4.8 degradation following the 6/5 multi-model error-rate incident.

Claude Code 2.1.167 — official release notes list only "bug fixes and reliability improvements" with no itemized details. Released again the same day after 2.1.166 (00:55 UTC, which carried detailed changes).

Claude Code 2.1.166 — New: a `fallbackModel` setting to configure up to three fallback models tried in order when the primary is overloaded or unavailable (`--fallback-model` now also applies to interactive sessions); glob support in the deny-rule tool-name position (`"*"` denies all tools), with allow rules rejecting non-MCP globs and unknown tool names in deny rules warned at startup. Improved: hardened cross-session messaging — messages relayed via `SendMessage` from other Claude sessions no longer carry user authority (receivers refuse relayed permission requests, and auto mode blocks them); `MAX_THINKING_TOKENS=0`, `--thinking disabled`, and the per-model thinking toggle now disable thinking on models that think by default via the Claude API (3P providers unchanged); Claude Code retries a turn once on the fallback model when the API rejects an unexpected non-retryable error (auth, rate-limit, request-size, and transport errors still surface immediately); `claude update` announces the target version before downloading instead of going silent; typing a URL into the `claude agents` list now filters to the session whose first prompt contained it. Fixed: a recurring "image could not be processed" error and extra token usage from an unprocessable image; remote sessions becoming permanently stuck after a brief backend disruption during worker registration at startup; flickering in JetBrains IDE terminals (IntelliJ, PyCharm, WebStorm, etc.) on 2026.1+ (now using synchronized output); Shift+non-ASCII characters (e.g. Shift+ä → Ä) dropped in Kitty-keyboard-protocol terminals (WezTerm, Ghostty, kitty); PowerShell command validation occasionally hanging far past its time budget on Windows when a killed process's children held its output pipes; orphaned `claude --bg-pty-host` processes spinning at 100% CPU after the daemon dies while connected on macOS; voice mode requiring `/login` to clear a stale auth check after toggling `/voice`; managed settings with an invalid entry silently disabling enforcement of their remaining valid policies; managed-settings `allowedMcpServers`/`deniedMcpServers` predicates not matching when they use `${VAR}` references; background agent sessions that entered a git worktree crash-looping with "No conversation found" when reopened from `claude agents`; duplicated thinking text in the Ctrl+O transcript view while streaming; `/doctor` showing a contradictory failed "Not inside a remote session" check when run inside a remote session; the cursor sticking at the end of the first line when typing a multiline prompt in the `claude agents` dispatch and reply inputs; blank lines between background agent rows in the task list on terminals without Unicode support.

Cybernews reports Anthropic is investigating customer-data-leak concerns following the 6/5 Claude outage — per the report, after the 6/5 Claude outage at ~15:08 PT (15:08 UTC, matching the timing of status.claude.com's same-day 15:19 UTC multi-model error-rate incident), a software engineer on X raised concerns that Claude may have returned responses meant for other users during the event, surfacing potential customer-data-leak concerns. An Anthropic spokesperson told Cybernews "we haven't seen any other reports or evidence of customer data being leaked but take any reports of this very seriously and are investigating," adding that the outage itself was "completely unrelated and would not have impacted customer data." Claude Code engineer Thariq Shihipar added that the incident was an infrastructure issue, not an inference bug. Community reactions were split — some users argued cross-user response leakage on the Claude platform was not isolated to this incident, while others thought it more likely Claude was simply hallucinating. The outage primarily affected Opus 4.7 and 4.8 models; as of publication the company had not confirmed any customer-data exposure.

Anthropic, via Claude Code lead Boris Cherny's (@bcherny) X account, announced a temporary doubling of Claude Cowork usage limits — Cherny posted on 2026-06-05: "We doubled Claude Cowork usage limits for the next month. This applies to your 5-hr rate limits. If you've been saving up a big messy project, now's the time." Claude Cowork's 5-hour rate limit is doubled "for the next month," encouraging users to push through large accumulated projects during the window. The announcement landed the same day as the status.claude.com 15:19 UTC multi-model error-rate incident, though the post did not explicitly tie the two together.

Anthropic published "Making Claude a chemist" — working with world-class synthetic, computational, and analytical chemists to evaluate Claude on nuclear magnetic resonance (NMR, Nuclear Magnetic Resonance) spectroscopy, one of a chemist's most common analytical inputs. Experimental design: 20 compounds drawn from synthetic chemistry preprints published after the models' training cutoff, comparing Claude Opus 4.7, Opus 4.6, and Sonnet 4.6 against dedicated NMR software ChemDraw and MestReNova. Key findings: Opus 4.7 matched the experimentally reported splitting pattern more often than any other tool; all three Claude models predicted sub-peak spacing to within 0.5 Hz roughly 80% of the time, far ahead of ChemDraw and MestReNova at 26–35%; on routine data prediction tasks Opus 4.7 — as a general-purpose model with no chemistry-specific fine-tuning — is now on par with or better than ChemDraw and MestReNova on average. Beyond prediction, Claude can also work the problem in reverse, proposing a structure directly from the high-resolution mass spectrum and 1D peak list a chemist would paste into a chat, with no additional setup. The significance for chemists: NMR interpretation (manually matching each peak in the spectrum to an atom in the proposed structure) is one of the most time-consuming steps in synthetic chemistry.

Elevated error rates across many Claude models — status.claude.com began investigating elevated error rates on many Claude models at 15:19 UTC on 2026-06-05, isolated the issue to Opus 4.7 and Opus 4.8 at 15:43 UTC (other models had recovered), continued mitigation on those two models at 16:46 UTC, returned success rates to expected levels and moved to monitoring at 17:13 UTC, saw all models fully recovered at 18:27 UTC, and marked the incident Resolved at 18:28 UTC, lasting ~3 hours 9 minutes (15:19→18:28); affected models spanned Opus 4.8, 4.7, 4.6, 4.5 and Sonnet 4.6 (per-model recovery: Opus 4.6 15:25, Sonnet 4.6 16:23, Opus 4.8 16:59, Opus 4.7 17:12, Opus 4.5 17:29 UTC), and affected components include claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork, with Major impact. Anthropic did not disclose a root cause. This is another in the same series following the 6/3 07:10 Opus 4.7/4.8 error-rate incident, reflecting recurring frontier-model error-rate incidents since the Opus 4.8 launch.

Claude Sonnet 4 (claude-sonnet-4-20250514) and Opus 4 (claude-opus-4-20250514) announced for deprecation, retiring from API on 2026-06-15; migration to Sonnet 4.6 and Opus 4.6 recommended.

Claude Code 2.1.165 — official release notes list only "bug fixes and reliability improvements" with no itemized details. Follows the 6/4 release 2.1.163 (2.1.164 was never published; the releases atom jumps straight from 2.1.163 to 2.1.165).

Claude Code 2.1.163 — New: `requiredMinimumVersion`/`requiredMaximumVersion` managed settings cause Claude Code to refuse to start when its version is outside the allowed range, directing the user to an approved version; new `/plugin list` command lists installed plugins (with `--enabled`/`--disabled` filters); `/btw` gains a "c to copy" shortcut that copies the raw markdown answer, preserving formatting when pasted elsewhere; Stop and SubagentStop hooks can now return `hookSpecificOutput.additionalContext` to give Claude feedback and keep the turn going without being labeled a hook error; Skills add `\$` escape syntax to include a literal `$` before a digit; stdio MCP servers now receive the same `CLAUDE_CODE_SESSION_ID` as hooks/Bash on `--resume`. Improvements: background agent sessions now update to a new Claude Code version in the background, so opening a session after an update no longer waits on a cold restart; clearer descriptions for built-in commands and skills in the `/` menu; the subscription-switch suggestion now shows in the startup announcement slot instead of a toast; `claude agents` dispatching from the state-grouped view now starts the session in the directory the agent view was opened from. Fixes: `claude -p` hanging forever after its final result when a backgrounded command never exits (background shells now stopped ~5s after the result once stdin closes); `claude -p` failing with "ANTHROPIC_API_KEY required" on Bedrock/Vertex/Foundry when `CI=true` and no Anthropic API key is set; bash commands failing under bazel and EDR-protected Go workflows (`$TMPDIR` was overridden to `/tmp/claude-{uid}` for all commands instead of only sandboxed ones, a regression in 2.1.154); Bash commands failing on Windows with "EEXIST: file already exists" on the session-env directory when read-only or inside OneDrive; org-managed permission rules not applying for the entire session when the managed settings fetch completed during startup on a fresh config directory; background sessions in `claude agents` losing running background tasks when reattached after an update; terminal misalignment and a multi-second hang when exiting the agent view via Esc; clicking Stop on a background-task chip not clearing it when the process was already gone; keyboard input becoming permanently unresponsive after a paste whose end marker is dropped by the terminal; hook `if: "Bash(...)"` conditions firing on every Bash command containing `$()` or `$VAR` (now also matches commands inside subshells and backticks); deny rules on home-directory paths (e.g. `Read(~/Desktop/**)`) not blocking Bash commands referencing the path via `$HOME`; a stray "(no content)" line left in the transcript after closing panel dialogs like /mcp and /plugins.

TechCrunch reports — At the Bloomberg Tech conference on 2026-06-04, Anthropic co-founder and President Daniela Amodei downplayed doubts about enterprise AI return on investment (ROI), arguing companies are still learning deployment strategies and that value will emerge "over time" as AI is woven into daily workflows. She named coding, financial services, legal, and healthcare as current primary value drivers, expecting expansion as familiarity grows. Revenue trajectory: annualized revenue jumped from roughly $9 billion at the end of 2025 to $47 billion by May 2026. On infrastructure, she said Anthropic—unlike rivals building proprietary data centers—prefers to manage demand prudently and avoid overextending on compute capacity. The remarks echo her same-session framing that training models is highly capital-intensive and public markets suit that need, against the backdrop of the 6/1 confidential S-1 submission and $965 billion valuation.

SiliconANGLE reported on Anthropic's call for a global pause on frontier AI development — unpacking the June 4 official post "When AI builds itself": head of internal research Marina Favaro and head of policy Jack Clark argue "it would be good for the world to have the option to slow or temporarily pause frontier AI development" so societal structures and alignment research can keep pace; Clark warned some models could reach recursive self-improvement capabilities within roughly two years. The piece adds competitive and skeptical context: the timing follows Anthropic's USD 65 billion raise and near-USD 1 trillion valuation, with the company preparing an IPO alongside OpenAI; analyst Rob Enderle called the move "more about strategic marketing than any concrete initiative" to slow development, and venture capitalist David Sacks accused Anthropic of pursuing a fear-based "regulatory capture" agenda to disadvantage open-source competitors.

Fortune reports — Anthropic's IPO will for the first time reveal the precise stakes and paper gains held by Amazon and Google. Amazon: cumulative investment of roughly $8 billion since 2023, with a paper value of about $74 billion as of April 2026; its stake is estimated in the mid-to-high teens (percent), and in Q1 2026 alone it booked $16.8 billion in pre-tax gains on its Anthropic position, added a further $5 billion after quarter-close, and committed up to $20 billion more — at a $965 billion valuation the position is worth roughly $135–160 billion. Google: court filings confirm about 14% in straight equity, hard-capped at 15%, worth roughly $135 billion at a $965 billion valuation, with up to $40 billion more committed ($10 billion immediate, $30 billion milestone-contingent); at a $1 trillion IPO valuation the stake could reach about $140 billion. The S-1 registration statement will disclose share count and price range and confirm the exact ownership percentages currently kept private.

Anthropic published "When AI builds itself" on the Anthropic Institute — head of internal research Marina Favaro and head of policy Jack Clark warn that AI is increasingly automating AI development itself, and that if current technical trajectories persist, systems could eventually design and train their own successors without human intervention (recursive self-improvement). Evidence: as of May 2026, more than 80% of the code merged into Anthropic's codebase was authored by Claude (in the low single-digit percentages before Claude Code launched in early 2025); engineers merged 8x as much code per day in Q2 2026 as in 2024; by April 2026 Claude reached a ~52x speedup on code-optimization tasks (vs. 4x in four to eight hours for skilled humans); AI reliability on long-horizon tasks roughly doubled every four months, progressing from 4-minute to 12-hour tasks within a year. The key threshold: AI already excels at executing specified tasks, but humans still hold an edge in "choosing which problems matter" and research-direction judgment — closing that gap would enable true recursive self-improvement. Anthropic argues for the option of a globally coordinated pause or slowdown of frontier model development to give society and alignment research room to keep pace, and plans in the coming months to convene policymakers, researchers, civil society, and other AI companies for coordination and deliberation.

Bloomberg (via investing.com) reported that Anthropic co-founder and President Daniela Amodei said at the Bloomberg Tech conference on 2026-06-04 that training AI models is "a very capital-intensive business" and that public markets are "very well-suited to that," citing high compute costs as a key driver behind the company's push for an IPO. The remarks came days after Anthropic confidentially submitted its draft S-1 to the U.S. SEC on June 1. Scale cited: SpaceX supplies AI compute capacity of roughly 325,000 Nvidia chips at about USD 1.25 billion per month, or USD 15 billion annualized; the company closed a USD 65 billion round last week at a USD 965 billion valuation and is ahead of OpenAI in the race to list. Amodei declined to comment on timing competition with OpenAI, noting only that the filing preserves the option to go public after the SEC review.

Claude Code 2.1.162 — Features & improvements: `claude agents --json` now includes `waitingFor` showing what a waiting session is blocked on (e.g. a permission prompt); explicitly listing Grep/Glob in `--tools` now provides the dedicated search tools on native builds with embedded search (previously silently ignored); `/effort` now confirms when your chosen level will persist as the default for new sessions; clicking a slash command in autocomplete now fills it into the prompt instead of running it immediately (press Enter to run); Remote Control now shows as a persistent footer pill (with a session link) instead of a startup message; Windsurf renamed to Devin Desktop in `/ide`, `/terminal-setup`, and `/scroll-speed` following the editor's rebrand; quieter startup (notices grouped by severity, session info and announcements share one line per launch), shorter and clearer startup warnings each with a concrete fix, and launch-prompt warnings now pinned below the input until acted on; failed turns show a compact warning line instead of a multi-line red error block; background service startup and `claude update` verification now wait out endpoint-security scanning of new binaries instead of failing after 5 seconds; removed the 'Claude in Chrome enabled' and 'marketplace installed' startup messages. Fixes: silent startup hang when the config directory is read-only or unwritable (now starts with in-memory config and surfaces errors instead of a blank screen); WebFetch permission rules not applied to built-in preapproved domains (explicit `WebFetch(domain:...)` deny/ask/allow now takes precedence); Windows permission rules never matching with backslash (`~\`, `\\server\share`) or case-variant paths, and Read deny rules not hiding files from Glob/Grep; an Esc interrupt sent at the very start of a turn silently dropped in stream-json/SDK sessions; API 400 `no low surrogate in string` for classifier side-queries and MCP server descriptions with emoji near a truncation boundary; MCP per-server `timeout` below 1000ms floored to a 1-second watchdog that aborted every tool call (sub-1000ms now ignored); the LSP `workspaceSymbol` operation returning no results (now accepts a `query` parameter); `claude agents` cutting live status text at 60-120 columns and session names at 40 columns (now grows with terminal width); `claude agents` attach occasionally bouncing back to the session list on the first try after a background-service restart; Ctrl+V image paste doing nothing in `claude agents`; backgrounding a session with the left-arrow silently losing the conversation when the background service can't start (now kept as a wakeable failed row); replies from the agents view that fail to send being lost (now queued for next session start); cross-session messaging (`SendMessage`) silently breaking when `CLAUDE_CODE_TMPDIR` or `$TMPDIR` points at a deep directory; opening a running background session from `claude agents` stalling for 5 seconds.

Anthropic published "How Anthropic enables self-service data analytics with Claude" on the claude.com blog — an internal deployment case: 95% of business analytics queries at Anthropic are now automated via Claude, with aggregate accuracy around 95%. Key finding: without skills, Claude's analytics accuracy didn't exceed 21%; adding skills pushes aggregate consistently above 95%, regularly reaching 99% in certain domains. The post also publishes two offline evaluation methods — (a) dashboard evals auto-generated by Claude and human-validated, and (b) long-tail evals where Claude generates plausible questions across domains. The data science team can now focus on more strategic work as a result.

Bloomberg (via Yahoo Finance) reports Anthropic has selected the lead banks for its IPO — Morgan Stanley and Goldman Sachs will lead the offering, with JPMorgan Chase also participating, and more banks may join later. This follows Anthropic's confidential S-1 draft filing with the SEC on June 1; the report says the IPO could launch as soon as October 2026, the company carries a $965 billion valuation from its most recent funding round (among the world's largest private companies), and the offering details may still change as discussions continue. Anthropic is racing OpenAI, which is also preparing to go public, toward the public markets.

TechCrunch reports that Lovable has signed a multi-year deal with Google Cloud expanding its usage roughly 5×, including expanded access to both Anthropic's Claude (widely used for coding tasks) and Google's Gemini models. Background: Google invested $10 billion in Anthropic in cash and compute credits in April and pledged an additional $30 billion contingent on Anthropic hitting performance milestones; Lovable's deal is a downstream materialization of that expanded credit pact — Lovable doubling down on both Claude and Gemini reflects a dual-model supply pattern in the vibe-coding / AI-coding-tool market.

Anthropic published "Lessons from building Claude Code: How we use skills" on the claude.com blog — the Claude Code team shares how they build and use skills internally. Key points: skills are "folders of instructions, scripts, and resources" that agents can discover and invoke to do things more accurately and efficiently, not merely markdown files; the team identifies nine categories of skills used inside Anthropic (from library references to infrastructure operations). Six core practices: (1) Don't state the obvious — focus on information that pushes Claude beyond its default behavior; (2) Build a Gotchas section — the highest-signal content in any skill is the Gotchas capturing Claude's common failure points; (3) Use file-system structure for progressive disclosure — organized folders with references/scripts/examples instead of one monolithic document; (4) Give Claude flexibility — provide necessary information but let the model adapt rather than over-specifying procedures; (5) Leverage persistent memory — store data in text logs or JSON via the `${CLAUDE_PLUGIN_DATA}` environment variable so Claude can reference its own history; (6) Include utility scripts — prebuilt code libraries let Claude focus on composition and decisions rather than reconstructing boilerplate. The team also tracks skill usage via PreToolUse hooks to measure effectiveness and identify popular versus underperforming skills.

Anthropic published "Running an AI-native engineering org" on the claude.com blog — Fiona Fung, Director of Engineering for Claude Code and Claude Cowork, shares how existing processes broke and were rewritten once agentic coding became the team's default way of working. Highlights: (1) Planning — shifted from six-month roadmaps (which were out of date by month three given Claude Code's pace) to "just-in-time planning," writing fewer design docs and discussing in PRs and prototypes instead; (2) Code review — moved to a trust-but-verify model where Claude handles style/linting, PR feedback, and catching and fixing bugs, while humans focus on domain expertise like security, legal, and product decisions; (3) Team structure — roles blurred (PMs now code, engineers do design and content), prioritizing "creative builders with product sense" and deep systems expertise over raw throughput; managers start as ICs, teams stay flat, and outdated processes get killed. Metrics tracked: engineers ship real code within their first week, PR cycle time decreased, and Claude-assisted commits are now the default (Fung notes no non-Claude commits on her team in four months).

Anthropic published "Introducing the Services Track and Partner Hub of the Claude Partner Network" — adding a "Services Track" tiering for services partners and a "Partner Hub" portal. The Services Track grades consulting/SI partners by certified practitioners, production deployments, and customer endorsements into three tiers: Select (10+ certified, 2 production customers, 1 public story), Preferred (100+ certified, 15 customers, 3 stories), and Global Premier (1,000+ certified, 100+ customers across 3+ regions, 15 stories); promotions are assessed each Jan 1 and Jul 1 (with an extra Oct 1 review in 2026). The Partner Hub lets partners track their standing against tier requirements in real time and lets customers find qualified firms; entry is free, with tiered discounts on certifications. Scale: 40,000+ firms applied and 10,000+ consultants earned Claude certifications, with major partners including Accenture (30,000 trained), Cognizant (350,000), Deloitte (470,000), KPMG (276,000+), Infosys, and PwC; backed by a $100M investment in partner training, technical support, and shared marketing.

Anthropic published "What we learned mapping a year's worth of AI-enabled cyber threats" — its Frontier Red Team analyzed 832 banned accounts tied to malicious cyber activity from March 2025 to March 2026 and drew three conclusions: (1) greater danger — attackers increasingly use AI in later, complex attack stages rather than for initial access; (2) greater autonomy — AI chains attack stages together with minimal human intervention, making traditional risk assessment unreliable; (3) framework gaps — MITRE ATT&CK does not capture AI-driven orchestration behaviors. Statistics: 67.3% (560 accounts) used AI for malware development and 6.5% (54) for lateral movement; 56% of accounts were medium-risk or higher in period two versus 33% in period one (a 1.7x increase); AI use for account discovery rose 8.9% while phishing fell 8.6%. Anthropic also published an interactive visualization on its Frontier Red Team blog, released an ATT&CK Navigator tool for defenders, and contributed the results to Verizon's 2026 Data Breach Investigations Report (DBIR).

Claude Opus 4.7 elevated error rate — status.claude.com began investigating elevated request error rates on Opus 4.7 at 07:10 UTC on 2026-06-03, marked a fix implemented and entered monitoring at 07:28 UTC, and resolved at 07:38 UTC, lasting about 28 minutes; impact spanned claude.ai, Claude Console (platform.claude.com), the Claude API (api.anthropic.com), and Claude Code; Anthropic did not disclose a root cause. It is yet another Opus 4.7 error-rate incident since Opus 4.8 launched, and the second platform incident of the day following the 04:17 UTC Claude Code service degradation.

Claude Code cloud services partially degraded — at 2026-06-03 04:17 UTC, status.claude.com began investigating an issue degrading some Claude Code services, including security reviews, code reviews, routines, and some Claude Code web sessions; the incident was marked resolved at 07:36 UTC, lasting about 3 hours 19 minutes. Anthropic has not disclosed a root cause.

Anthropic announced via the official @ClaudeDevs X account an update to Claude Code's `/fork` behavior — `/fork` now spins up a background agent with your exact context (system prompt, tools, conversation history, model) plus prompt cache, and the result is returned to your current session; the old `/fork` behavior has been renamed to `/branch`, which still copies the transcript to a new session for you to drive yourself. The change is not listed in v2.1.161's release notes (shipped the same day at 21:58 UTC) — it is a companion behavior adjustment announced via X.

Claude Code 2.1.161 — New: `OTEL_RESOURCE_ATTRIBUTES` values are included in metric datapoint labels, enabling usage slicing by custom dimensions like team/repo; the `claude agents` list shows `done/total` before details when a task expands, with peek showing the longest-running item; `/mcp` collapses never-logged-in claude.ai connectors into a "Show unused connectors" row; a failed Bash command in parallel tool calls no longer cancels other calls in the same batch — each tool returns its result independently; the fullscreen-mode Linux clipboard now uses whichever of `wl-copy`/`xclip`/`xsel` is available, writing to both clipboard and PRIMARY selection to support middle-click paste, and the "hold {key} for native selection" hint shows the correct key per terminal; VSCode adds a suggestion to disable terminal GPU acceleration (or run `/terminal-setup`) to fix garbled text. Fixes: `/effort` dialog, workflow animations, and prompt keyword shimmer not respecting "Reduce motion"; `forceLoginOrgUUID`/`forceLoginMethod` managed settings wrongly blocking third-party provider sessions (Bedrock, Vertex, Foundry, Mantle; regression since v2.1.146); background subagent output polluting `claude -p` stdout (`--output-format text` or `json`); `/usage-credits` triggering re-login for Team and Enterprise admins instead of pointing to the org usage settings page; `/autofix-pr` falsely reporting "cannot run on the default branch" inside git worktrees or other repos; the `--resume` picker not listing current-directory sessions when the directory is not a git worktree (e.g. jj workspaces); Windows hooks explicitly invoking bash (e.g. `/usr/bin/bash script.sh`) failing ("command not found" or "cannot execute binary file"); OpenTelemetry log events (`user_prompt`, `api_request`, `tool_result`, `tool_decision`) silently dropped before telemetry initialization completes; `claude mcp` list/get/add printing secrets to the terminal (`${VAR}` no longer expanded; credential headers and URL secrets now masked); Workflow agents dispatched with `isolation: "worktree"` from background sessions being blocked from editing files in their own worktree; background sessions dispatched by `claude agents` starting on the daemon environment's stale model instead of the `settings.json` setting; a possible crash rendering Write tool results after resuming a session; the screen stuck on running when a subagent's finalize stage errors; `EADDRINUSE` when `CLAUDE_CODE_TMPDIR` is set to a deep path and some tools bind Unix sockets under `$TMPDIR`. Improvements: stabilized layout-engine JIT compilation profile for better terminal rendering performance; rendering performance for large file writes.

Storyboard18 reports the root cause and quota fallout of the major 2026-06-02 Claude outage — the outage originated in Claude Code's sub-agent system: a malfunction reportedly caused sub-agents to multiply excessively, driving an abnormal surge in token consumption. Many users (especially on Pro and Max plans) experienced a “silent quota drain” — Claude still appeared responsive, but usage limits ran out within minutes even under normal use. Anthropic deployed an emergency quota reset for affected Pro and Max subscribers to restore their limits while engineers fixed the issue. The disruption affected Claude Code, Cowork, claude.ai and the API; customers accessing Claude through Google Cloud Vertex AI or Amazon Bedrock were not affected. This is the media-side root-cause analysis of the same multi-model error-rate incident on status.claude.com that ran 06:04–11:49 UTC (which disclosed no official root cause); the sub-agent multiplication account is media-reported and not officially confirmed by Anthropic.

TechCrunch reports Anthropic is scaling Claude Mythos to critical infrastructure across 15+ countries — filling in details the official “Expanding Project Glasswing” post left unnamed: roughly 150 new organizations spanning Australia, Canada, France, Germany, Italy, Switzerland, the Netherlands, Spain, Belgium, Sweden, India, Japan, New Zealand and South Korea; confirmed organizations gaining access include Okta, Samsung, SK Hynix, SK Telecom, NATO and the EU cybersecurity agency ENISA. New sectors focus on power, water, healthcare, communications and hardware — critical-infrastructure areas underrepresented in the initial rollout. TechCrunch adds competitive context: Anthropic expects other AI companies to soon develop models as capable as Mythos Preview, and OpenAI has already released its own cybersecurity model GPT-5.5-Cyber; an attack on these partners’ codebases “could affect more than 100 million people,” with global and national-security stakes.

Anthropic published “Expanding Project Glasswing” — expanding its cybersecurity program: building on roughly 50 initial partners that began using Claude Mythos Preview in early April to scan codebases for vulnerabilities (finding more than 10,000 high- or critical-severity flaws to date), it is adding about 150 more organizations across more than 15 countries. Newly included sectors that were previously underrepresented are power, water, healthcare, communications and hardware; many new partners are vendors whose codebases are relied upon by many other organizations, and for most partners “a major attack could affect more than 100 million people.” New partners must meet security requirements before gaining model access. Anthropic says it is working as quickly as it can to safely release Mythos-level capabilities to general access once stronger safeguards are in place, but gives no specific timeline.

The National reports Claude suffered a major service outage on 2026-06-02 — users could log in but, about three minutes after submitting a prompt, received “A bit longer, thanks for your patience” followed by “Due to unexpected capacity constraints, Claude is unable to respond to your message. Please try again soon”; Anthropic said it was aware of the issue and working on a fix. The disruption affected the main Claude app, Claude Console, Claude API and Claude Code. This is the media-side coverage of the same multi-model error-rate incident on status.claude.com that ran 06:04–11:49 UTC (about 5 hours 45 minutes); the status page disclosed no official root cause. The coverage comes right after Anthropic’s confidential S-1 IPO filing and $965 billion valuation.

City A.M. reports — UK accountancy firm Grant Thornton UK will roll out Anthropic Claude to all partners and employees across audit, tax, advisory and support functions, phased in between June and August 2026, as part of a £500m programme aimed at shifting work from process-heavy routine tasks toward higher-value judgement and advice. Partners will be required to use Claude for tasks such as analysis, drafting and synthesizing large volumes of information, with training and governance frameworks provided; the move forms part of the firm’s newly launched GT Augment AI governance framework, and it plans to open a Digital Experience Centre in London later this year. CEO Malcolm Gomersall said “clients don’t pay for process; they pay for judgement,” and that AI “will make client outcomes stronger by helping us deliver more insight, challenge and advice,” stressing clients still work directly with advisers who remain accountable for the advice given.

Axios (reporter Madison Mills, via Yahoo Finance) reports Anthropic faces an enterprise AI-spending backlash ahead of its IPO — enterprise customers, Anthropic's main revenue source, are entering an "AI sticker shock" phase, and any pullback in AI spending could weaken its revenue right before listing. Evidence: a Bain survey of ~1,000 companies found 40% reported AI cost savings below 10%; an early Anthropic investor warned companies are "waking up to how much they're spending on Claude"; an AI consultant said a CFO client "accidentally spent half a billion dollars on Claude in a single month" with no usage caps set. OpenAI CEO Sam Altman called corporate AI-cost concern "the most fair criticism of AI so far." The piece also notes Anthropic surpassed OpenAI in business customers in April 2026 and is tracking toward nearly $50B in annual revenue and a first profitable quarter — but enterprise spending decisions will determine whether AI labs' IPO valuations hold.

Claude Developer Platform official release note (2026-05-12): Fast mode (research preview) now supports the Claude Opus 4.7 model. By setting `speed: "fast"`, `model: "claude-opus-4-7"`, and the `fast-mode-2026-02-01` beta header in an API request, developers get significantly faster output token generation (up to roughly 2.5x output tokens/sec), priced at 6x standard Opus rates across the full context window (including requests over 200k input tokens). Pricing, rate limits, and access are the same as for Opus 4.6 fast mode; interested customers must join the waitlist. Fast mode does not support the Batch API and is not available for Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, or Microsoft Azure Foundry — it is offered only via the direct Anthropic API (with the beta header) and supported native integrations (Claude Code, Cursor, Windsurf, v0, Warp). Fast mode first launched for Opus 4.6 in February 2026.

Elevated errors across multiple Claude models — status.claude.com began investigating elevated request error rates across multiple models at 06:04 UTC on 2026-06-02, marked the issue identified with a fix being implemented at 06:39 UTC, deployed a fix and entered monitoring at 10:42 UTC, and marked it resolved at 11:49 UTC, lasting about 5 hours 45 minutes; affected scope covers claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com) and Claude Code; Anthropic has not disclosed a root cause. It follows the five model error-rate incidents on 6/1 and is the first on 6/2, spanning multiple models.

Claude Code 2.1.160 — Security hardening: added a prompt before writing to shell startup files (`.zshenv`, `.zlogin`, `.bash_login`) and `~/.config/git/`, which could otherwise lead to unintended command execution; `acceptEdits` mode now prompts before writing build-tool config files that grant code execution (`.npmrc`, `.yarnrc*`, `bunfig.toml`, `.bazelrc`, `.pre-commit-config.yaml`, `.devcontainer/`, etc.). Edit no longer requires a separate Read after viewing a file with `grep` — single-file `grep`/`egrep`/`fgrep` commands now satisfy the read-before-edit check. Fixes: copy-on-select not writing to the Windows clipboard on WSL (now uses PowerShell interop instead of OSC 52, unsupported by terminals like MobaXterm); restoring a completed session from `claude agents` dropping chat history and re-running the original prompt; background sessions re-attached after overnight retire losing their conversation and re-running the original prompt; `claude --bg` occasionally failing with "socket missing" during cold-start on a loaded machine; on Windows the directory a background session started in being undeletable after `claude rm` until the daemon exits; background agents that resumed work shown under Completed; `claude agents` freezing for seconds when returning to the session list due to the auto-updater re-checking on every exit; Esc, arrow keys and typing becoming unresponsive on Windows under heavy CPU load when attached to a background session or in agent view; background agents emitting terminal sync-output markers to unsupporting terminals (Apple Terminal, tmux), causing render artifacts; mouse wheel scrolling prompt history instead of the transcript right after opening a session from the agents list; CJK IME composition appearing at the bottom-left instead of at the input caret in the `claude agents` view; valid `file:///C:/...` links rewritten to a broken path on Windows terminals; voice mode failing to connect when the project directory or branch name contains non-ASCII or special characters; the auto mode unavailability message on third-party providers (Bedrock/Vertex/Foundry) pointing to the model instead of the `CLAUDE_CODE_ENABLE_AUTO_MODE` opt-in; `/effort ultracode` blaming the dynamic workflows setting when the model cannot run xhigh (ultracode is no longer offered on unsupported models); model-not-found errors suggesting `--model` when running via the SDK or other hosts; Claude's past replies disappearing from scrollback when resuming a brief mode session with brief mode off; vim mode `p` pasting on the line below instead of at the cursor when yanked with `v$`. Improvements: performance opening recently-inactive background agent sessions; auto mode classifier latency (less reasoning on routine actions, fewer "could not evaluate this action" blocks); background-session teardown (`claude rm`/`stop`, idle reap) now sends SIGTERM to running shell subprocesses before SIGKILL so cleanup handlers run. Removed `CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE` (now a no-op) and the JetBrains plugin install suggestion at startup. Renamed the dynamic-workflow trigger keyword from `workflow` to `ultracode` (the word "workflow" no longer triggers a run; asking in your own words still works; the keyword is highlighted in violet in the prompt input).

Claude Sonnet 4.6 degraded performance (2nd Sonnet 4.6 incident on Jun 1) — status.claude.com began investigating elevated errors on Claude Sonnet 4.6 requests at 18:28 UTC on 2026-06-01 and marked it resolved at 19:06 UTC, lasting about 38 minutes; impact spanned claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause. This was the fifth model error-rate incident of the day (Opus 4.7 twice at 06:48 and 12:15, Sonnet 4.5 once at 09:08, Sonnet 4.6 twice at 14:40 and 18:28).

Elevated errors on Claude Sonnet 4.6 — status.claude.com identified the issue and began deploying a fix at 14:40 UTC on 2026-06-01, entered monitoring at 15:21 UTC, and marked it resolved at 15:41 UTC, lasting about an hour; the impact spanned claude.ai, Claude Console (platform.claude.com), the Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause. This was the fourth model error-rate incident on 2026-06-01 (two Opus 4.7 incidents at 06:48 and 12:15, one Sonnet 4.5 at 09:08) and the first to affect the Sonnet 4.6 model.

Anthropic confidentially submitted a draft S-1 registration statement to the U.S. Securities and Exchange Commission (SEC), kicking off its initial public offering (IPO) process at a valuation that could approach $1 trillion — just after closing a $65 billion Series H on 5/28 at a $965 billion post-money valuation (surpassing rival OpenAI's $852 billion March valuation). Anthropic said it has not yet set the number of shares or the price, and that IPO timing will depend on market conditions and other factors. Its run-rate revenue has crossed $47 billion (up from about $9 billion at the end of 2025); the Wall Street Journal previously reported it expects a roughly 130% revenue surge toward its first operating-profit quarter. OpenAI is also preparing an IPO, with the two AI labs racing toward a Wall Street debut.

Claude Opus 4.7 elevated errors — status.claude.com began investigating at 12:15 UTC on 2026-06-01, identified the root cause and began a fix at 12:19 UTC, deployed the fix into monitoring at 12:48 UTC, and marked it resolved at 13:33 UTC, lasting about 1 hour 18 minutes; the impact spanned claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork. This is the seventh Opus 4.7 error incident since the Opus 4.8 launch.

Claude Sonnet 4.5 elevated errors — status.claude.com began investigating at 09:08 UTC on 2026-06-01, deployed a fix into monitoring at 09:13 UTC, and marked it resolved at 09:28 UTC, lasting about 20 minutes; the impact spanned claude.ai, the Claude Console (platform.claude.com), the Claude API (api.anthropic.com), and Claude Code.

Anthropic published the official announcement “Anthropic confidentially submits draft S-1 to the SEC” — confirming that Anthropic, PBC has confidentially submitted a draft registration statement on Form S-1 to the U.S. Securities and Exchange Commission for a proposed initial public offering, giving it the option to go public after the SEC completes its review, subject to market conditions and other factors. Anthropic states the number of shares and the price have not yet been set; the notice is issued under Rule 135 of the Securities Act of 1933 and is neither an offer to sell nor a solicitation of an offer to buy. This is the T1 primary source underlying same-day media reports (post-money valuation of $965 billion, revenue run-rate above $47 billion) from TechCrunch and others.

At Snowflake Summit 26, Snowflake announced a deepened strategic partnership with Anthropic — expanding Claude's integration across Snowflake Cortex AI to meet enterprise demand for governed, production-ready AI. Highlights: (1) Snowflake Cortex Code (Snowflake's fastest-growing product ever, now surpassing 7,100 users) is powered by leading models like Claude and purpose-built for Snowflake schemas; (2) Snowflake Intelligence, a personal agent for knowledge workers, is powered by models like Claude and enables natural-language queries and reasoning across enterprise data; (3) the Cortex Agents framework supports use cases including customer-support automation, data analysis, and core operations; (4) Snowflake is one of six launch partners in the Claude Marketplace, letting customers apply existing Anthropic commitments toward Snowflake AI capabilities to consolidate AI spend. Claude operates directly on data within Snowflake, so customers can pick the model best suited to their workload without moving sensitive data outside the Snowflake environment. Adopting enterprises include Basis, Block, Carvana, eSentire, Indeed, and Notion. This builds on the companies' $200 million expanded partnership from December 2025.

Claude Opus 4.7 elevated errors — status.claude.com began investigating at 06:48 UTC on 2026-06-01, deployed a fix into monitoring at 07:08 UTC, and marked it resolved at 07:22 UTC, lasting about 34 minutes; affected components spanned claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), and Claude Code; Anthropic disclosed no root cause. This is the sixth Opus 4.7 error-rate incident since 5/25 (5/25, two on 5/27, 5/28, 5/30, 6/1), still occurring after Opus 4.8 launched on 5/28.

Claude Code 2.1.159 — internal infrastructure improvements only (no user-facing new features, fixes, or behavior changes).