Archive — 2026-07Week 1

Elevated errors for some models (resolved) — primarily affecting Claude Fable 5 — status.claude.com logged an "Elevated errors for some models" incident affecting claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation of elevated error rates on requests to some models began 2026-07-07 20:02 UTC, success rates returned to normal by 20:28 UTC with work continuing to prevent recurrence, a 21:47 UTC update said the issue primarily affected Claude Fable 5, monitoring began at 23:19 UTC, and the incident was declared resolved at 23:46 UTC — about 3 hours 44 minutes in total. It was the second platform incident of 7/7 (following the early-morning Sonnet 5 errors incident), coinciding with the transition window in which Fable 5 moved off subscription plans to usage-credit billing after 7/7 as originally scheduled.

Claude Cowork expands to iOS/Android/web — beta rolls out to Max plans first, doubled Cowork usage limits extended through August 5 — On 2026-07-07, Anthropic announced on the claude.com blog ("Claude Cowork on web and mobile: hand off work anywhere") that Claude Cowork (the desktop Claude Code–style agent first shipped in January) now extends to the web and iOS/Android. Users can hand off a task from a desktop, monitor it on a phone, and pick up the finished output on any device; even with the laptop closed, Claude keeps working in the cloud, scheduled tasks can run with no device online, and Claude sends a notification asking for permission when it needs one. Web and desktop combine chat and Cowork into a single home tab (one sidebar, one search, one place for Projects and Artifacts) — no more separate tabs. The beta rolls out over the next several weeks starting with the Max plan, with more plans to follow; to support the expansion, Cowork usage limits are doubled through 2026-08-05. On the web, Cowork opens from the claude.ai home screen; on mobile, from the sidebar of the Claude iOS/Android app.

VentureBeat: Anthropic first-time discloses Claude Cowork's usage breakdown — more than 90% is not software development, business operations tops the list at 33.4% — Alongside the Claude Cowork mobile/web expansion on 2026-07-07, VentureBeat published Anthropic's first public breakdown of Cowork usage. Drawing on 1.2 million anonymized Cowork sessions from more than 600,000 organizations sampled 2026-05-11 to 05-31: business process operations leads at 33.4% — pulling scattered updates into reports, building onboarding checklists, reconciling spreadsheets, etc. — common in finance, HR, and administration; content creation and copywriting comes in second at 16.4% — drafts, decks, social posts, proposals, and communications work — largely from marketing and management roles; software development accounts for just 8.7%. Anthropic frames business operations and content creation as "the work around the work" — connective work that spans every role in an organization but rarely appears in anyone's core job description. The breakdown runs against the media's long-standing framing of Cowork as a coding-focused product and is seen as key evidence for expanding Cowork's positioning to general knowledge workers.

Bloomberg: Microsoft gradually replaces OpenAI and Anthropic models in Excel and Outlook with in-house MAI models — to cut Copilot costs — In an exclusive report on 2026-07-07, Bloomberg says Microsoft, seeking to reduce AI costs, has begun replacing OpenAI and Anthropic models with in-house MAI models across software including Excel and Outlook — in these two widely used spreadsheet and email applications, tens of thousands of AI prompts each week are now completed by internally built MAI models, whereas the apps previously leaned more heavily on outside models. Strategic motivation: Microsoft AI chief Mustafa Suleyman said in June, "We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost." Timeline context: at Microsoft's Build developer conference in June, it announced seven new AI models at once, including one it claims can match — at lower cost — the coding ability of the still-popular previous-generation Anthropic Opus 4.6 model. The report is the first public disclosure of the scale at which Microsoft's MAI models are being used inside Copilot productivity apps, suggesting a structural contraction of Anthropic's model shipments to Microsoft, one of its largest customers.

Claude adds Trusted Devices for Remote Control on Team and Enterprise plans — Anthropic announced in the Claude release notes (2026-06-25) that admins on Team and Enterprise plans can now require members to verify their device before remotely viewing or steering a local Claude Code session, adding device-level security verification to remote control access.

Claude's Microsoft 365 connector gains write tools — draft and send email, manage calendars, create and update OneDrive/SharePoint files — Claude's official release notes record on 2026-07-07 that the Microsoft 365 connector expands beyond read-only search to write operations: with write tools enabled, Claude can draft, send, and organize email, manage calendar events, update mailbox settings, and create and update files in OneDrive and SharePoint, always within each member's existing Microsoft 365 permissions; read and search capabilities remain unchanged and Teams stays read-only. A Microsoft Entra administrator must consent to the updated permission set and an organization admin must enable the tools before members can use them. Released the same day (7/7) as Claude Cowork's expansion to web and mobile.

Anthropic brings Claude Code and Claude Cowork to government — FedRAMP High environment, public beta open now — Anthropic announced on the claude.com blog on 2026-07-07 that it is bringing Claude Code and Claude Cowork to government: public-sector teams can use Claude Code to build and modernize the software systems behind public services, while agency staff can delegate memo drafting, RFP review, casework, and presentation building to Cowork. Inference runs on FedRAMP High authorized infrastructure and conversation logs live only on agency-managed devices; compliance features include hash-chained audit logs covering all administrative actions, two-person approval for sensitive operations on Anthropic's side, usage exports containing only metering data to protect sensitive content, a published FedRAMP Secure Configuration Guide, penetration-test summaries available under NDA, and deployment via standard agency MDM platforms. Billing supports appropriation-bound standard seats or custom seat tiers with spend/model caps, fixed incremental purchases, and hard no-overage limits; the admin console provides per-user/per-model usage tracking and automatic balance alerts. Anthropic is the sole contracting and billing party, so agencies need no separate cloud-provider relationship; the company stresses that "agencies will get new capabilities at the same cadence as commercial users." The public beta is effective immediately, with new customers applying at claude.com/solutions/government.

Anthropic published the official data report “How people are using Claude Cowork” — analyzing 1.2 million anonymized, aggregated Claude Cowork sessions from 2026-05-11 to 05-31 across more than 600,000 organizations, classified into 20 work categories by an automated system with no individual sessions reviewed by humans. Business process and operations leads at 33.4%, content creation and copywriting 16.4%, software development only 8.7%, DevOps and infrastructure 7%, research and intelligence 6.4%, data analysis and BI 5.8%. About half of all usage is “the work around the work” — tasks that span many roles but aren't core responsibilities; the top two categories make up roughly half of usage and are overwhelmingly connective in nature (assembling and structuring information rather than specialized domain expertise). This is the official report cited by VentureBeat's same-day coverage.

TechCrunch: "the coding agent wars are spilling into the rest of the office" — Cowork on mobile and web in a race for workspace-platform dominance — TechCrunch reported on 2026-07-07 on Claude Cowork's expansion to web and mobile, noting that since the January desktop launch it is now available across devices for Max subscribers: start a task at your desk, get progress notifications on your phone, and pick up the output later (even with the laptop closed). The report cites Anthropic's analysis of 1.2 million anonymized sessions from over 600,000 organizations in late May: business-process automation leads at 33.4%, content creation takes 16.4%, and software development accounts for just 8.7% — evidence Cowork's use has moved far beyond engineering scenarios; Anthropic says "our goal is for it to be the reference point for how people think about integrating AI products into their daily workflows." On the competitive front, OpenAI is pushing Codex beyond a developer tool toward reports, spreadsheets, and data analysis, and the piece frames the contest as being about controlling the workspace platform, not merely offering a stronger chatbot.

ECB tells 110 directly supervised banks to submit AI cybersecurity action plans by end of October — names Anthropic's Mythos model as reshaping the threat landscape — Euronews reported on 2026-07-07 (with Bloomberg publishing the same day) that Claudia Buch, chair of the ECB's Supervisory Board, has written to the CEOs of the 110 banks it directly supervises (including Deutsche Bank, BNP Paribas, and Santander), requiring action plans against AI-driven cyber threats by 2026-10-31, covering vulnerability management (faster software patching), AI-powered cyber defense, and third-party vendor oversight, with infrastructure modernization over the longer term. The letter calls frontier AI models "a long-term shift in the threat landscape rather than a temporary phenomenon" that "significantly amplifies" the speed and scale at which risks emerge; the report singles out Anthropic's Mythos model as especially adept at identifying computer-system vulnerabilities, noting Anthropic initially restricted the full version over misuse concerns before releasing a public version with safety guardrails. The European Systemic Risk Board (ESRB) has raised its systemic cyber-risk assessment to "severe" and warns of EU dependence on non-European AI developers. After the deadline the ECB will analyze plans bank by bank, discuss them with each institution, and run a horizontal analysis to identify common weaknesses and best practices across the sector.

IBTimes: Chinese AI models are winning U.S. customers on price — a new cost war for OpenAI and Anthropic — IBTimes reported on 2026-07-07 (CNBC published a same-day report on the same trend) that Chinese models have taken over 30% of tokens processed weekly on OpenRouter since February 8 (versus an 11% average over the prior 12 months and 4.5% in H1 2025), peaking at 46%. On cost, OpenRouter's Justin Summerville says leading Chinese open-source models cost 60–90% less than comparable Anthropic and OpenAI offerings; Zhipu's GLM 5.2 scored within one percentage point of the Opus 4.8 model on a closely watched agentic benchmark at roughly one-fifth the cost, and performs competitively on certain cybersecurity benchmarks. On adoption, Vercel's Harpreet Arora reported GLM 5.2's daily token volume grew about 27x in its first week after launch; AI startup Lindy has moved all its traffic from Claude to DeepSeek, with CEO Flo Crivello saying the switch will save millions of dollars within months. Brookings' Kyle Chan notes companies have shifted from adopting AI "regardless of which model" to scrutinizing operating costs as deployments scale. The report also notes Anthropic has confidentially filed for an IPO while expanding government partnerships.

Elevated errors on Claude Sonnet 5 (resolved) — status.claude.com logged an "Elevated errors on Claude Sonnet 5" incident affecting claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation began 2026-07-07 06:58 UTC and the incident was declared resolved at 07:37 UTC, lasting about 39 minutes. It marks a second consecutive day of platform incidents following the two on 7/6 (elevated errors on several models, and on Claude.ai); Sonnet 5 launched on 6/30 as the default model for Free and Pro plans.

Alibaba confirms ban on Anthropic models — effective July 10, staff switch to in-house Qoder, citing back-door security concerns — BeInCrypto reported on 2026-07-07 (republished by Yahoo Finance; CNBC ran the same story that day) that Alibaba has told employees to stop using Anthropic's AI tools for work from 2026-07-10, remove Anthropic models from work computers, and switch to Qoder, its in-house coding assistant. The stated reason is "back-door security risks" — days before the ban, developers found Claude Code inspecting users' time zones and proxy data and allegedly inserting subtle markers into prompts sent to Anthropic's servers. Context: in June Anthropic wrote to the U.S. Senate Banking Committee accusing Alibaba of the "largest known distillation attack" against Claude; TechCrunch first reported the ban rumor on July 4, and this report confirms the effective date and internal replacement. The report also ties the move to China's Qinglang campaign tightening domestic AI oversight (over 14,000 AI products removed and 26,000 accounts suspended recently).

Claude Code v2.1.202 released — New: a "Dynamic workflow size" setting in /config controls how large Claude generally makes dynamic workflows (small/medium/large agent counts; an advisory guideline, not an enforced cap); telemetry emitted by workflow-spawned agents adds workflow.run_id and workflow.name OpenTelemetry attributes so a workflow run's activity can be reconstructed from OTel data. Fixes: a crash in the inline Ctrl+R history search when accepting or cancelling while still scanning the history file; /rename on background sessions being reverted when the job restarts, breaking addressing by the new name; transient mTLS handshake failures when settings were re-applied during an in-place client certificate rotation; commands sent from Remote Control (mobile/web) into an interactive session failing with "Unknown command"; images and files sent from Remote Control without a caption being silently dropped; the sign-in URL printed by claude auth login and claude mcp login --no-browser not being reliably clickable when it wraps over SSH (now emitted as a single hyperlink); opening a chat from claude agents sometimes failing with "currently running as a background agent" followed by a worker crash/respawn loop; workflow scripts with unicode quote escapes in strings being corrupted before parsing (parse errors now show the offending line instead of always blaming TypeScript); voice dictation retrying in an unbounded loop when the microphone or audio recorder fails (repeated capture failures now pause voice input); /remote-control sessions showing the wrong permission mode in the mobile and web apps; resuming a session by name, or opening the resume picker, taking minutes and using a large amount of memory in repositories with many git worktrees; installer and updater downloads failing immediately with "aborted" when a proxy or network drops the connection mid-download (transient drops now retry); re-invoking an already-loaded skill appending a duplicate copy of its instructions to context. Improvements: /workflows agent list layout (wider titles, a dedicated time column, shorter model names, no per-row tool-call counts); clearer MCP error when a server config has url but no type (suggests "type": "http" instead of the misleading "command: expected string"); /review <pr> is back to a fast single-pass review — use /code-review <level> <pr#> for the multi-agent review at a chosen effort level.

Claude Tag elevated GitHub operation failures (resolved) — status.claude.com logs an incident "Claude Tag seeing elevated GitHub operation failures" affecting Claude Code: investigation began 2026-07-06 21:49 UTC, with Anthropic investigating elevated error rates on calls made to Claude Tag and users possibly seeing errors such as "GitHub operations denied". On 07-07 15:30 UTC the team identified the issue and adjusted the load sent to GitHub to reduce the error rate; at 15:52 UTC, after observing a lower error rate, the incident was declared resolved with continued monitoring — lasting about 18 hours, the longest-running of the recent platform incidents. Claude Tag is Anthropic's virtual-employee tool launched in late June that works across organizations within Slack.

Elevated errors on Claude.ai (resolved) — also broke OAuth log-in for Claude Code — status.claude.com logs an "Elevated errors on Claude.ai" incident affecting Claude.ai and Claude Code OAuth log-in: investigation began 2026-07-06 19:18 UTC; at 19:28 UTC the issue was identified, with the added impact that users could not log in to Claude Code via Claude.ai OAuth; a fix was applied at 19:46 UTC with recovery of Claude.ai and related services observed; the incident was declared resolved at 20:05 UTC, lasting about 47 minutes. This is separate from the multi-model elevated-errors incident earlier the same day (starting 07:36 UTC), making it the second platform incident of July 6.

Fable 5's removal from subscription plans is temporary — Claude Code lead engineer Thariq Shihipar clarifies on X it will be restored "as soon as capacity allows" — Addressing concerns that Claude Fable 5 becoming a paid add-on is permanent after it leaves Pro/Max/Team and select Enterprise plans following 2026-07-07 (switching to usage-credit billing at $10 per million input tokens and $50 per million output tokens, roughly double Opus 4.8), Claude Code lead engineer Thariq Shihipar (@trq212) responded on X on 2026-07-06: while Fable 5 comes off subscriptions after July 7, "we aim to restore Fable as a standard part of our subscriptions as soon as capacity allows, as we mentioned in our original blog post." Anthropic attributes the arrangement to extremely high and unpredictable demand following the model's reinstatement, framing it as a temporary measure rather than a permanent pricing change.

Anthropic publishes "A global workspace in language models" — J-Space discovered inside Claude — On 2026-07-06 Anthropic posted a new research paper in its official research section introducing a novel interpretability tool called the Jacobian lens (J-lens), which uncovered a small but important "privileged workspace" inside Claude that the company dubs J-space. The J-lens works by computing the average mathematical effect of a given internal activation pattern on the model's future output of specific tokens. J-space has three key properties: (1) it is tiny — holding only "a few dozen" concepts at a time and accounting for less than a tenth of the model's overall activity, yet carrying most of the work that matters for safety; (2) it is observable and interveneable — Claude can report the contents of J-space when asked and can modulate them on request, and researchers can causally alter Claude's outputs and task performance by intervening directly in J-space; (3) it emerges from training — the activation patterns were not hand-designed but arose spontaneously during training. The authors draw a parallel to Bernard Baars' neuroscience global-workspace theory — the brain as a theater with many parallel processors working backstage while only a small "spotlight" of information gets broadcast to the whole theater and becomes conscious content. Anthropic emphasizes that the research "does not claim Claude is conscious" and "does not claim Claude has subjective experience," warns against reading it as "AI is alive" clickbait, and positions J-space as a key entry point for future detection of misalignment and scheming.

Illinois Governor Pritzker signs SB 315, the AI Safety Measures Act — the first US state law to require annual third-party safety audits of frontier labs; Anthropic publicly voices support — Illinois Governor JB Pritzker signed Senate Bill 315 (Artificial Intelligence Safety Measures Act) on the morning of Monday, 2026-07-06 in Chicago, establishing what the state calls "the nation's strongest" AI safety, transparency, and accountability framework. It applies only to "large frontier" developers — companies with annual gross revenue over $500 million that train models using compute exceeding 10^26 operations — effectively covering roughly a dozen firms including OpenAI, Anthropic, Google DeepMind, Meta, and xAI. Key provisions: mandatory annual independent third-party safety audits (a first in any US state AI law), required publication of safety frameworks and transparent reporting, and protections for employees who report safety concerns. Passed with bipartisan support; effective 2027-01-01. Industry response: Anthropic said the law "takes the safety practices leading labs already follow voluntarily — publishing a safety framework, transparent reporting, protecting whistleblowers — and helps establish a baseline that every leading AI developer is expected to meet"; OpenAI also publicly supported it. It marks a significant positive stance by Anthropic on US state-level AI regulation.

Anthropic signs 20-year, $19 billion Kentucky data-center lease with TeraWulf — TeraWulf announced on 2026-07-06 that Anthropic is leasing approximately 401 MW (critical IT load) at its Justified Data Campus in Hawesville, Kentucky under a 20-year lease expected to generate about $19 billion of contracted revenue over the initial term (a simple average of roughly $950 million per year, with actual rent depending on phased delivery timing and lease escalation mechanics). Initial capacity is expected to come online in the second half of 2027, with the full 401 MW ramp completed by early 2028. In a separate announcement the same day, TeraWulf agreed to sell its 50.1% interest in the Abernathy Joint Venture (a 168 MW AI data-center campus in Abernathy, Texas, established in 2025) to an investor group led by Fluidstack for $530 million paid in three installments ($250 million within 14 days of signing, $150 million by 2026-12-31, and about $130 million subject to adjustments by 2027-04-30) — a premium exit over TeraWulf's roughly $450 million invested capital, with proceeds redirected to wholly owned AI infrastructure. TeraWulf shares rallied sharply on 7/6 as the market validated its pivot from crypto mining to AI data centers, and the deal marks another large compute contract for Anthropic on top of its $1.25-billion-per-month SpaceX arrangement.

Washington Post deep dive on the covert US-China AI war — the full inside story of Anthropic's hidden tracking code and the Alibaba distillation allegations — The Washington Post published a national-security deep dive on 2026-07-06 (by Cate Cadell, Gerrit De Vynck and Nitasha Tiku; syndicated on Yahoo News) tracing the covert battle between Anthropic and Chinese AI firms: starting in March 2026, Anthropic deployed hidden tracking code in Claude Code that checked whether a user's computer was set to Chinese time zones and used web domains linked to certain Chinese AI companies, aiming to unmask rivals suspected of hijacking its technology; after a developer exposed the mechanism and privacy advocates criticized Anthropic for surveilling its own users, the company removed it last week. The report also cites Anthropic's letter to US senators alleging that Alibaba's Qwen team used roughly 25,000 fraudulent accounts to generate more than 28.8 million exchanges with Claude for model distillation. On the competitive front, Chinese models have matched US capabilities within months for over a year: a recent assessment by security firm Semgrep found Zhipu AI's free model outperformed the Claude Opus 4.8 model at identifying computer vulnerabilities — "They're very close... and they cost you nothing," said Securin CEO Srinivas Mukkamala — while Chinese users circumvent restrictions via proxy networks and resellers, accessing US AI services for roughly $1 a month, with premium subscriptions around $12.

Anthropic publishes official feature "The Making of Claude Code" — the inside story of Claude Code's journey from internal CLI to coding agent. Published 2026-07-06 in the features section of anthropic.com, it is told by the researchers, engineers, and early users who built Claude Code, recounting how it grew from an internal CLI tool into Anthropic's coding agent. The page offers two reading modes — "Read in Terminal" and "Read as Article" — the former rendered in terminal style as an interactive reading experience, a nod to the product's CLI origins.

Media coverage of Anthropic's "J-space" discovery — an emergent, consciousness-like global workspace inside Claude, with attention centering on the AI-consciousness debate — Crypto Briefing reports (2026-07-06) that the J-space is an emergent activation structure inside Claude models that functions like the brain's global workspace for sharing information across cognitive processes: "Claude can report the contents of its J-space when asked. It can modulate those contents upon request." On safety, monitoring J-space activity could uncover hidden motivations in model behavior and detect prompt-injection attacks. The report stresses that Anthropic explicitly does not claim Claude is conscious or has subjective experience — describing information only as "consciously accessible" — and situates the research in Anthropic's year-long trajectory: the October 2025 introspective-awareness work and the April 2025 model-welfare initiative. Axios noted the same day that the paper uses the word "conscious" more than 200 times, focusing media attention on the AI-consciousness debate.

Anthropic publishes Government of Alberta case study: using Claude Code to scan and fix cybersecurity vulnerabilities across government systems at scale — On 2026-07-06, Anthropic published the official case study "Government of Alberta uses Claude to find and fix cybersecurity vulnerabilities across government systems." Alberta's Ministry of Technology and Innovation deployed Claude Code (using both Opus and Sonnet models) across all 27 provincial ministries, covering 1,280 applications and 3,400 code repositories holding sensitive data such as tax records and social services case files. Assessment phase: about 50 Claude agents autonomously scanned 466 million lines of code in parallel, finishing in just 20 hours — a task estimated to take 6.5 years by traditional methods; a two-stage process first flagged known vulnerability patterns with a rules engine, then reviewed findings to pinpoint exact files and line numbers for developer verification. Remediation phase: Claude Code generated fixes, created automated tests where needed, and even rebuilt legacy systems in modern languages — a Java subsidy portal that originally took five months to build was reconstructed in four to five days. Ongoing protection: dedicated Claude agents monitor throughout development, including "red team" agents that simulate attacks and "blue team" agents that assess defenses against international security standards. Minister of Technology and Innovation Nate Glubish said: "By using AI to find and fix vulnerabilities across our systems, we accomplished in hours what would have taken a traditional approach years to complete."

Authors who opted out of Anthropic's $1.5 billion book-piracy settlement file a new $75 million copyright suit — On 2026-07-06, Yahoo Finance and other outlets reported that a group of authors who opted out of the $1.5 billion Bartz class-action settlement have brought a separate copyright lawsuit against Anthropic seeking $75 million, alleging Anthropic sourced their works from shadow libraries such as LibGen and used them to train Claude without licensing or payment. The suit builds on the California court's prior finding that downloading pirated copies is a separate act of infringement, while training on legally obtained books qualifies as fair use. Plaintiffs seek statutory damages of up to $150,000 per willfully infringed work; by contrast the earlier $1.5 billion settlement covered roughly 500,000 pirated books at about $3,000 per author, and the opt-out strategy lets individual authors pursue far larger per-work damages. (The report does not name the plaintiffs, the exact filing date, or the court.)

MacRumors covers the July 6 multi-model elevated-error incident — mainstream consumer tech media attention on Claude's recurring outages — MacRumors reported on the 2026-07-06 Claude disruption: the status page flagged elevated errors across several models at 07:36 UTC, spreading to Fable 5 by 08:45 UTC; four of six monitored services (claude.ai, the Claude API, Claude Code, and Claude Cowork) showed degraded performance while Claude Console and Claude for Government were unaffected; resolved at 09:13 UTC, about 1.5 hours end to end. The article places the incident in the context of Anthropic's recent pattern of outages — previous disruptions were likewise attributed to demand outpacing available compute capacity (driven by enterprise adoption of Claude Code and a surge in consumer signups) rather than technical failures.

Elevated errors across several models (resolved) — status.claude.com logged an "Elevated errors on several models" incident affecting claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation began at 07:36 UTC on 2026-07-06 (initially elevated errors on Claude Haiku 4.5), the issue was identified at 08:45 UTC as Fable 5 also began returning errors, and it was declared resolved at 09:13 UTC, lasting about 1 hour 37 minutes. It continues a recent pattern of recurring error-rate spikes across frontier and lightweight models.

Alibaba bans employees from using Claude Code, effective 2026-07-10, and pushes its in-house Qoder instead — On 2026-07-04, citing reports from Reuters, Morningstar, and Reddit discussions, TechCrunch reported that Alibaba has classified Anthropic's Claude Code as "high-risk software" and, on security grounds, will bar all employees from using it starting July 10, directing them to the group's own Qoder tool. The security concern stems from researchers finding that Claude Code, since March 2026 (media point to around v2.1.91 shipped in April), contained a covert detection mechanism that could secretly identify whether a user was in China or affiliated with a Chinese AI lab — matching The Register report already logged here on 7/1 (Anthropic merged a pull request removing the covert code on 7/1). Anthropic's Thariq Shihipar confirmed the feature was "an experiment launched in March" meant to prevent unauthorized-reseller account abuse and model distillation, adding the team has "landed stronger mitigations since then." The ban extends the two companies' standoff — Anthropic had accused Alibaba of an "industrial-scale distillation attack" using roughly 25,000 fake accounts and over 28.8 million interactions against Claude from 2026-04-22 to 06-05, and has blocked Chinese firms' access. SCMP, Cybernews, Korea's Seoul Economic Daily, and others covered it in parallel.

Claude Code v2.1.201 released — the changelog lists a single change: Claude Sonnet 5 sessions no longer use the mid-conversation system role to deliver harness reminders.

Elevated errors across many models incident (resolved) — status.claude.com logged an incident affecting claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation began 2026-07-03 23:42 UTC, a fix was implemented and monitoring began 2026-07-04 00:19 UTC, and it was marked resolved at 00:20 UTC, lasting about 38 minutes.

Elevated errors on Claude Fable 5 and Mythos 5 incident (resolved) — status.claude.com logged an incident affecting claude.ai, the Claude Console (platform.claude.com), the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation began 2026-07-03 21:54 UTC, the issue was identified and remediation started 21:55 UTC, monitoring began 22:00 UTC, and it was resolved at 22:06 UTC, lasting about 12 minutes. It continues the pattern of recurring frontier-model error rates following Fable 5's global restoration on July 1.

Claude Code v2.1.200 released — Changed: AskUserQuestion dialogs no longer auto-continue by default (opt into an idle timeout via /config); the "default" permission mode is renamed "Manual" across the CLI, --help, VS Code, and JetBrains (--permission-mode manual and "defaultMode": "manual" are accepted alongside the old default). Fixed: startup crash when disabledMcpServers/enabledMcpServers in .claude.json is set to a non-array value; background sessions silently stopping mid-turn after sleep/wake or when reopening a stalled session; background sessions re-running a turn cancelled with Esc after a stall respawn; background agents never restarting after a crash left a stale daemon.lock whose PID the OS reused; daemon handover now judging build recency by the version's embedded build timestamp so a reinstalled older build can't take over; background-agent roster issues (transient corruption permanently disabling orphan cleanup, older binaries not preserving fields written by newer versions, socket auth tokens stripped during daemon restarts); subagents cut off by a rate limit before producing any text returning an empty result instead of failing cleanly; control bytes from background-agent output reaching the terminal in the agent view; claude agents --plugin-dir <dir> not showing the plugin's agents and skills when the flag follows agents; project-scoped plugins not loading from git worktrees of the same repo; the /mcp server list not tracking focus for screen readers and magnifiers; voice dictation showing a misleading "Voice connection failed" when a recording captures no audio; rendering flicker under tmux 3.4+ (now enabling synchronized terminal output). Improved: screen-reader output (decorative glyphs hidden, transcript symbols read as short labels, nested tables read as "Header: value." lines); the install script now explains when installation is killed by the system running out of memory.

Financial Times: Anthropic steps up efforts to block Chinese firms from accessing Claude via offshore workarounds — In a July 3, 2026 scoop, the FT reported Anthropic has intensified detection and blocking of the methods Chinese companies use to bypass its terms of service and reach Claude: overseas subsidiaries (e.g., Ant Group via a Singapore entity), personal subscriptions with VPNs (ByteDance reportedly reimbursed engineers), "transfer station" relay services routing queries through foreign accounts, and cloud infrastructure such as Microsoft Azure. Detection signals include monitoring account computer time zones and usage patterns to flag accounts acting as relays for China-linked firms. The move extends 2025 terms barring companies more than 50% owned by entities in unsupported regions (China, Russia, Iran, North Korea), which Anthropic said were "intended to prevent workarounds." The FT noted these practices don't violate US or Chinese law but breach Anthropic's terms. CEO Dario Amodei disclosed in February 2026 that Anthropic had "forgone several hundred million dollars in revenue" by restricting CCP-linked access and blocking CCP-sponsored cyberattacks. Anthropic frames the policy as reducing legal, regulatory, and security risks and preserving a "democratic advantage" in frontier AI.

Claude Code v2.1.199 released — New: stacked slash-skill invocations like `/skill-a /skill-b do XYZ` now load all leading skills (up to 5), not just the first; transient server rate-limit errors (429s unrelated to your usage limit) are now retried automatically with backoff for subscribers instead of failing the turn; `CLAUDE_CODE_RETRY_WATCHDOG` now raises the default retry count for non-capacity transient errors to 300 and lifts the cap of 15 on `CLAUDE_CODE_MAX_RETRIES`. Fixes: SSL certificate errors (TLS-inspecting proxies, missing `NODE_EXTRA_CA_CERTS`, expired certs) burning retries before showing actionable guidance (they now fail immediately with the fix hint); streaming responses being discarded when the API emits a mid-stream overloaded/server error after partial output (the partial is now kept with an incomplete-response notice); subagents cut off by a rate limit or server error silently failing instead of returning their partial work to the parent; subagents reporting API errors (e.g. usage limit reached) as successful results; the background-agent daemon on Linux killing itself and every running agent every ~50 seconds after an unclean shutdown left a corrupted worker record; background agents failing to cold-start over SSH on macOS with "Could not switch to audit session" (regression in 2.1.196); `claude stop` being silently undone when it raced a background-agent respawn; background job progress indicators stalling for minutes during long commands; background sessions on memory-starved machines showing a generic error (they now indicate low memory and suggest freeing resources); remote sessions briefly flapping between Working and Idle in the agent view when a background agent completes; idle subagents vanishing from the agent panel while others were still working (surplus idle agents now collapse into an expandable summary row); typing `/model` or `/fast` while viewing a subagent silently opening the lead's model picker (a notice now explains the command applies to the lead); `SessionStart`, `Setup`, and `SubagentStart` hooks silently hiding stderr when exiting with code 2 (the error now shows in the transcript); `claude --dangerously-skip-permissions daemon <subcommand>` being treated as a chat prompt; `SendMessage` silently misrouting when a re-spawned agent reuses a previous agent's name (the tool now detects the mismatch and asks the caller to retarget); opening or resuming a session with no new messages needlessly growing the transcript file; backgrounding a session with `←` or `/background` dropping its `/color` from the agent view row; resetting a corrupted config file from the startup recovery dialog destroying it unrecoverably (it now backs up the file first); Claude in Chrome repeatedly opening the reconnect page when sessions run from different builds or config directories; plan mode not prompting for state-changing browser tool calls (read-only `browser_batch` calls are now correctly auto-allowed). Improvements: `claude agents` session rows show pull-request links as bare `#N` without the redundant "PR" label.

TechCrunch/The Information: Anthropic in talks with Samsung to build its first custom AI chip, eyeing 2nm process and advanced packaging — On 2026-07-02, TechCrunch, citing an exclusive from The Information the same day, reported that Anthropic has been in contact with Samsung Electronics to explore having Samsung manufacture its first custom AI chip. Anthropic is evaluating Samsung's 2-nanometer (2nm) process and advanced packaging facilities, but the project remains at an early stage — no decisions yet on what the chip will be used for, how it will fit into a server, or how powerful it will be, and Anthropic could still walk away. Anthropic recently hired Clive Chan, an early member of OpenAI's custom-chip team, seen as a signal of its silicon push. In its statement, Anthropic said that "a diversified hardware stack — including chips from Google, Amazon, and Nvidia — will continue to be pivotal to its compute strategy," with AWS Trainium, Google TPUs, and Nvidia GPUs remaining central. Samsung is one of the investors in Anthropic's $65B Series H in May, positioning it uniquely as both investor and potential manufacturing partner. The move comes just after OpenAI's custom-chip partnership with Broadcom, highlighting the industry trend of frontier AI companies accelerating a move away from Nvidia dependence toward in-house silicon; a Reuters report in April had noted Anthropic was mulling its own chip in response to supply constraints — this is the first concrete external collaboration signal.

Fortune: Fable 5/Mythos 5 are restored, but U.S. AI policy is still a mess — a de facto licensing regime denied officially, and trust damage already done. Fortune AI editor Jeremy Kahn analyzed the post-export-control landscape on 2026-07-02, arguing the U.S. is running "what is essentially a licensing regime for frontier AI models, while officially denying that this is the case," with rules seemingly improvised by officials, opaque and inconsistent. Second, the damage is done — even with controls lifted, enterprises now recognize that relying on American AI models poses strategic risks and are exploring fallback options, especially open-weight alternatives. Third, the open-source security threat is rising: Chinese open-weight models are becoming dangerously capable, and guardrails in open models can easily be stripped away — "if an attacker has access to a model's weights, there is always a jailbreak that can be found." Looking ahead, the U.S. is reportedly developing voluntary cybersecurity standards, and Anthropic is working with major companies on a jailbreak-risk assessment framework (the CJS severity framework published the same day).

Anthropic details Fable 5's cyber safeguards and formally proposes the CJS jailbreak-severity framework — On 2026-07-02 Anthropic published “More details on Fable 5's cyber safeguards and our jailbreak framework,” expanding on the protections previewed in the 6/30 Fable 5 redeployment post. First, safety classifiers sort cyber activity into four tiers: prohibited use (ransomware, malware development, defense evasion, data exfiltration — always blocked); high-risk dual use (legitimate security work such as penetration testing and exploit development — blocked until access controls for known good actors are in place; ICS/SCADA, telecom-core SS7/Diameter, and financial-infrastructure assessments also fall here); low-risk dual use (OSINT and vulnerability identification on par with existing tools — mostly allowed, some blocked as a safety margin); and benign use (patch management, incident response, secure coding — allowed with minimal false positives). Fable 5's safety margin is larger than prior models'. Second, together with Glasswing partners including Amazon, Microsoft, and Google, Anthropic formalized the Cyber Jailbreak Severity (CJS) framework: four scoring axes — capability gain (0–4), breadth of gain (0–2), ease of weaponization (0–2), discoverability (0–2) — map to five levels from CJS-0 (informational) to CJS-4 (critical, 9–10), with logarithmic escalation between levels. Anthropic stresses it does not block all vulnerability discovery (“responsibly disclosing a newly discovered vulnerability is clearly in the national interest”) and accepts jailbreak reports via cyber-safeguards@anthropic.com and a HackerOne bounty program.

Claude Enterprise adds admin analytics, model entitlements, and spend alerts — Three updates: (1) the admin analytics dashboard now shows usage and cost by group and by user, alongside outputs like artifacts created, files edited, and skills and connectors used; Claude Code gets dedicated tabs for active developers, session counts, and estimated productivity, and "analytics chat" answers natural-language queries (e.g. "Which teams doubled their Claude usage?") with exportable charts. (2) Model defaults and entitlements: admins can set the default model for new conversations across chat, Cowork, and Claude Code, and control which models are available to specific roles or the whole org, so routine work does not default to the most expensive option. (3) Spend alerts: admins are notified at 75% / 90% of the org spend limit; users get in-app notifications at 75% / 95% and can request an increase directly from their admin. The Admin API automates cost-control workflows, and the Analytics API integrates with Datadog Cloud Cost Management and CloudZero. Available on Claude Enterprise plans.

Elevated error rate on Claude Opus 4.8 (resolved) — status.claude.com logged an incident affecting claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation began 2026-07-02 00:38 UTC, the issue was identified and a fix implemented at 00:58 UTC, and it was declared resolved at 01:19 UTC, lasting about 41 minutes. This continues the pattern of recurring frontier-model error-rate incidents since Opus 4.8 launched — Opus 4.8 error-rate incidents also occurred on 6/27 and 6/30.

Claude Code v2.1.198 released — Features: Claude in Chrome is now generally available (GA); `claude agents` adds background agent notifications (sessions that need input or finish fire the `Notification` hook, `agent_needs_input`/`agent_completed`); new `/dataviz` skill for chart and dashboard design guidance with a runnable color-palette validator; Gateway adds Claude Platform on AWS (`anthropicAws`) as an upstream provider, and model-not-found responses now advance the failover chain; background agents launched from `claude agents` now commit, push, and open a draft PR after finishing code work in a worktree instead of stopping to ask; the built-in Explore agent now inherits the main session's model (capped at opus) instead of haiku; subagents and context compaction inherit the session's extended thinking configuration to improve delegated-task quality. Fixes: transient network drops mid-response aborting the turn (ECONNRESET and similar now retry with backoff), excessive background classifier requests from sandboxed processes repeatedly hitting the same host, background tasks stuck on "Running" in web/desktop/VS Code panels after finishing or resuming, agent teams (a teammate dying on an API error now reports "failed" and messaging a stuck teammate wakes it to retry), the `/diff` panel not refreshing on branch switch or out-of-session commit, markdown tables overflowing in fullscreen, AWS/Mantle sessions dead-ending on "Please run /login" when the STS token expires (`awsAuthRefresh` now auto-runs), "no route to host" for local-network hosts in macOS background agents (Local Network entitlements), `claude --bg` combined with `--print`/`-p` silently creating an unattachable session (conflicting flags now rejected up front), `.claude/rules/` conditional rules not loading via symlinked paths, and plan mode not auto-allowing read-only tool calls at startup. Improvements: syntax highlighting upgraded to highlight.js 11, removed the `/agents` wizard (ask Claude or edit `.claude/agents/` directly), and better API retry UX (error reason shown after the second attempt, status-page link replacing the spinner tip when overloaded).

Fable 5/Mythos 5 suspension incident marked resolved on the status page — ~18-day duration — status.claude.com logged an incident titled "We've suspended access to Claude Mythos 5 and Claude Fable 5" affecting claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: monitoring began 2026-06-13 00:50 UTC (details at anthropic.com/news/fable-mythos-access) and it was marked resolved 2026-07-01 19:26 UTC (details at anthropic.com/news/redeploying-fable-5), lasting about 18 days. This is the platform status record for the full window — from the US export-control directive requiring both frontier models be disabled for all foreign persons starting 6/12, through the Commerce Department's full lifting on 6/30, to Fable 5's global restoration on 7/1 — matching the suspend-to-restore arc described in Anthropic's official "Redeploying Fable 5" post and broad media coverage.

UC Berkeley Computer Science Division chair Jelani Nelson takes leave to join Anthropic — Theoretical computer scientist Jelani Nelson announced on X on 2026-07-01 that he has joined Anthropic as a Member of Technical Staff, taking a leave of absence from UC Berkeley (retaining his tenured position): "Excited to work with many talented, mission-driven people on the defining technology of our time." Nelson has chaired the Computer Science Division of UC Berkeley's EECS department since 2019; an MIT graduate who previously taught at Harvard, he specializes in streaming algorithms and data sketching, and Anthropic recruited him to strengthen the mathematical rigor of data-processing efficiency for large language models. Nelson also founded AddisCoder, a summer coding program in Addis Ababa, Ethiopia that has sent disadvantaged students to MIT, Harvard, and other elite institutions, and received the ACM Eugene L. Lawler Award for humanitarian contributions. Media framed the move as part of a wave of elite academics pausing tenure for frontier AI — extending Anthropic's pre-IPO talent offensive following Nobel laureate John Jumper and Google Gemini core researchers Jonas Adler and Alexander Pritzel.

Fortune: Restored access to Fable 5/Mythos 5 signals a "necessary truce" between Anthropic and the US government — a strategic pivot under IPO pressure — Fortune analyzed on 2026-07-01 the lifting of export controls and full restoration of both frontier models, noting the two-plus-week standoff ended with Anthropic committing to closer government collaboration: Commerce Secretary Howard Lutnick said the government spent those weeks ensuring the models had "alignment" with US interests; Anthropic pledged to "scale up our government collaboration" and to establish industry standards with Amazon, Microsoft, and Google using common scoring systems for assessing model vulnerabilities. The piece highlights the business incentive — Anthropic, valued at roughly $965 billion, has confidentially filed for an IPO and is racing OpenAI to go public, and "open hostility with the Trump administration could stand in the way of Anthropic's IPO plans," suggesting accommodation of government concerns was driven by listing considerations. It also recaps the standoff's origin: the controls only banned foreign nationals, but because they applied so broadly, Anthropic suspended access for all users including its own foreign employees; officials feared hostile actors could jailbreak the models and bypass built-in safety guardrails.

Korea JoongAng Daily: The lifting of export controls is asymmetric — Fable 5 opens globally while Mythos 5 stays limited to vetted institutions, with Korea likely to resume access via Project Glasswing — Korea JoongAng Daily reported on 2026-07-01 that although the US Department of Commerce has lifted export controls on Anthropic's two frontier models, Claude Fable 5 and Mythos 5, the two differ in availability: Fable 5 is the public-tier model made broadly available and is restored globally, while Mythos 5 is the same underlying model but with fewer guardrails and, owing to its higher-risk capabilities in areas such as cybersecurity, remains limited to vetted institutions. The ban took effect June 12, was partially eased June 26, and was fully lifted on Tuesday (June 30) after roughly two and a half weeks of suspension, with Anthropic promising to begin restoring access the next day. Regionally, the Korean government and firms have signed onto Project Glasswing — the program through which Anthropic provides Mythos 5 to a small set of vetted organizations — but actual deployment had stalled under the US restrictions; with the controls gone, Mythos 5 access is now likely to resume in Korea through approved partners, though Anthropic has not confirmed when or for whom. Commerce Secretary Howard Lutnick said Anthropic has agreed to proactively detect and address security risks in the models and committed to working with the government on protocols and standards.

The Register: Anthropic removes covert code in Claude Code that detected Chinese competitors — Developers found that since March 2025 Claude Code contained a hidden detection mechanism: it checked for base-URL environment-variable overrides, read the system timezone and hostname, compared hostnames against a list of known Chinese AI labs, account resellers, and gateway domains hidden behind XOR and base64, and used invisible Unicode markers to encode proxy classifications into seemingly plain-English system prompts — aimed at detecting unauthorized resellers and model distillation. After the allegations surfaced on Reddit with a corroborating verification report on GitHub, Anthropic merged a pull request removing the mechanism on July 1, saying it has implemented "stronger mitigations" without giving details.

9to5Mac: US lifting of Fable 5/Mythos 5 export controls was in fact a conditional release — Anthropic agreed to three obligations in exchange for waiving the export license — On 2026-06-30 at 5:24 PM PT (00:24 UTC on 07-01), 9to5Mac reported, citing a letter from US Commerce Secretary Howard Lutnick to Anthropic, that Commerce lifted export controls on the two frontier models Claude Fable 5 and Mythos 5 on the condition that Anthropic fulfill three commitments: proactively detect and address security risks in the models, work with the government on protocols for future releases, and report any malicious activity found in its own models. After meeting these conditions, Anthropic's products no longer require an export license. Lutnick said, "Over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5 to ensure alignment across the US Government and strengthen America's leadership in AI." Anthropic will begin restoring access to Fable 5 on Wednesday (7/1), though it did not specify whether restoration would be immediate or gradual. These conditions correct the framing of earlier reports that the lifting "appeared to have no attached conditions" — the release in fact came with security commitments.

Yahoo News/multiple outlets: US Commerce Department fully lifts export controls on Fable 5 and Mythos 5, ending a nearly three-week shutdown standoff — On 2026-06-30 CNBC, Axios, NBC and other outlets reported (carried by Yahoo News) that the US Department of Commerce lifted export controls on Anthropic's two most powerful models, Claude Fable 5 and Mythos 5, on Tuesday (6/30) evening, with access to be restored on Wednesday (7/1). Fable 5 had been forced offline on 6/12, just 3 days after its 6/9 launch, on national-security grounds (severe cybersecurity risk) — about three weeks in total. Commerce Secretary Howard Lutnick said his office had 'worked closely with Anthropic over the past two weeks to analyze and approve Fable 5 to ensure alignment across the US Government and strengthen America's leadership in AI.' After Commerce first cleared Mythos 5 on 6/26 for about 100 trusted organizations working on cybersecurity and critical infrastructure, both models are now fully approved for release. The lifting appears to carry no conditions; Anthropic also has built-in classifiers that route high-risk cybersecurity and biology queries to safer handling.

Anthropic officially announces the lifting of Fable 5/Mythos 5 export controls, Fable 5 returns globally on July 1, and proposes an industry framework for scoring jailbreak severity — Anthropic announced on 2026-06-30 in its official blog post 'Redeploying Fable 5' that the US Department of Commerce had that day lifted export controls on Claude Fable 5 and Mythos 5. Fable 5 returns starting the next day (Wednesday, July 1) to users globally on the Claude Platform, Claude.ai, Claude Code, and Cowork; through July 7 it is included at up to 50% of weekly usage limits for Pro/Max/Team and select Enterprise plans, then transitions to usage-credit billing. This is the full reversal—after about three weeks—of the June 12 national-security-based takedown triggered by an Amazon researcher's demonstration of a jailbreak technique (which could induce the model to reveal software vulnerabilities), following the June 26 initial clearance of Mythos 5 for over 100 trusted US institutions. Anthropic also proposed an industry-wide four-criterion framework for scoring jailbreak severity—capability gain (whether it exceeds existing tools), breadth of capability gain (how many distinct offensive tasks it enables), ease of weaponization (how much human effort to turn it into an attack), and discoverability (specialist knowledge vs. widely public)—developed with Amazon, Microsoft, Google, and Glasswing partners, aiming to become a common standard for assessing AI jailbreaks.

Claude Code v2.1.197 released — Shipped alongside the Claude Sonnet 5 launch: Sonnet 5 becomes the default model in Claude Code, with a native 1M-token context window and promotional pricing of $2/$10 per million input/output tokens (through 2026-08-31). Update to version 2.1.197 for access. This release's changelog lists only this single item.

Claude Sonnet 5 launched — On 2026-06-30 Anthropic introduced Claude Sonnet 5 (model id `claude-sonnet-5`) on its official blog, positioned as an agentic model whose performance is close to Opus 4.8 but at lower prices. It is now the default model for Free and Pro plans, and is available to Max, Team, and Enterprise users as well as in Claude Code and the Claude Platform. Introductory pricing (through 2026-08-31) is $2 per million input tokens and $10 per million output tokens; from 2026-09-01 it becomes $3 input and $15 output. Anthropic says Sonnet 5 is a substantial improvement over Sonnet 4.6 on agentic dimensions like reasoning, tool use, coding, and knowledge work; testers describe it as more agentic — finishing complex tasks where predecessors stopped short and checking its own output without being asked. On agentic search and computer use it delivers strong capability at lower cost, while Opus 4.8 offers greater accuracy at a higher price. Safety assessments found a lower rate of undesirable behaviors than Sonnet 4.6, making it generally safer in agentic contexts.

Claude Science workbench launched in beta — On 2026-06-30 Anthropic introduced Claude Science, an AI workbench for scientists that integrates the tools and packages researchers most commonly use into a single environment, enabling literature analysis, multi-step research, auditable artifacts, and flexible access to compute. It is pre-configured for genomics, single-cell, proteomics, structural biology, cheminformatics, and more, with over 60 curated skills and connectors backed by more than 60 scientific databases. It can build environments and manage compute on a laptop, a cluster, or on-demand GPUs, drafting a plan and asking before reaching new resources, with review or revoke available. The beta is on macOS and Linux (locally or remotely via an SSH/HPC login node) for Pro, Max, Team, and Enterprise plans. An accompanying AI for Science program funds up to 50 projects with up to $30,000 in credits each; applications close 2026-07-15, awards are announced 7/31, and projects run 9/1 to 12/1. Access at claude.com/science.