NEC and Anthropic Sign Strategic Collaboration — Claude Goes to 30,000 Employees and Inside BluStellar
NEC Corporation announced a strategic collaboration with Anthropic this morning Tokyo time, with three operative pieces. (1) Claude Opus 4.7 and Claude Code will be deployed to approximately 30,000 NEC Group employees globally to build what NEC is calling “one of Japan’s largest AI-native engineering teams.” (2) Claude is being embedded inside NEC BluStellar Scenario, the company’s flagship business-transformation consulting platform, starting with the “Data-Driven Management” and “Customer Experience Transformation” scenarios. (3) The two companies will jointly develop AI agents tuned to Japanese enterprise workflows, including agents wired into NEC’s industry verticals.
This is the biggest enterprise commitment Anthropic has booked in Asia and the most concrete deliverable from the regional push that started with the GIC-led $30B Series G. NEC’s 30,000-seat Claude Code rollout puts Claude on more developer desktops in Japan than any single competitor. The BluStellar piece is the more strategic of the two: NEC’s consulting practice does north of $4B annually and lands the company inside large bank, telco, and government modernization deals where the AI vendor decision typically gets made by the integrator. The agent-development work is the optionality — if Anthropic and NEC ship Japanese-vertical agents that work, that’s a model-distribution channel into the keiretsu network that Microsoft and Google have spent years trying to build.
Mythos Heads to European Banks Within Days — UK and EU Rollout Subject to Vendor Security Checks
Reuters and follow-on reporting through Yahoo Finance, US News, and FStech: Anthropic plans to widen Claude Mythos Preview availability to banks in Europe and the UK “within days,” with the rollout gated on additional security checks following Monday’s third-party vendor incident. Bank of England governor Andrew Bailey, ECB president Christine Lagarde, and Canada’s finance minister all flagged Mythos as a serious regulatory challenge at last week’s IMF spring meetings in Washington. German Bundesbank president Joachim Nagel went further, arguing publicly that all institutions should get access to preserve a level playing field.
Read the timing. The breach disclosure landed Monday night. The third-party vendor environment is what Anthropic is now hardening. The fact that the European bank rollout is still going forward this week — rather than getting paused for a 30-day audit — tells you the access-control story has not lost its shippable form. The bigger structural shift: Mythos started as a 40-organization defender preview and is now ramping toward what looks like every G7 systemically-important bank within thirty days. The “narrowly gated” framing is functionally over. The product is now “a vetted-but-broad defender deployment to critical-infrastructure financial institutions.” That’s a different kind of product, and it requires a different vendor management story than the one Anthropic shipped two weeks ago.
Australia Confirms Direct Mythos Working Relationship — Banking Association Says Banks Are Engaged With Regulators
A spokesperson for Australian Home Affairs Minister Tony Burke confirmed Thursday that the federal government is “actively working with software providers and Anthropic directly to track emerging vulnerabilities exposed by Mythos.” Simon Birmingham, CEO of the Australian Banking Association — representing the country’s major commercial lenders — said in a parallel statement that banks are engaging with regulators to ensure the financial system remains safe. The Reserve Bank of Australia and the Reserve Bank of New Zealand confirmed Wednesday they are monitoring the release and coordinating with regulators globally.
Pair it with the Japanese FSA’s high-level Friday meeting that brought MUFG, SMFG, Mizuho, the Bank of Japan, and the Tokyo Stock Exchange to the same table on Mythos. The picture across APAC: every major financial regulator is now in active coordination with Anthropic on a frontier model that has not formally launched. The implicit question for Anthropic’s policy team is whether this is a transitional state or the new equilibrium — do future model launches require this scale of pre-launch regulator engagement, or is Mythos a one-off because of the offensive-cyber capability profile? Answering wrong on either side is expensive.
Claude Code 2.1.118 — Vim Visual Mode, /usage Replaces /cost and /stats, Hooks Can Now Call MCP Tools Directly
Today’s build is a power-user release. Headline: full vim visual mode (v) and visual-line mode (V) with selection, operators (d, y, c, >, <), and proper visual feedback — the request from heavy keyboard-driven users that has been on the issue tracker for months. /cost and /stats are merged into /usage with the legacy commands kept as typing shortcuts that open the right tab; clean consolidation that removes the “which one tells me what” question. Themes get a real management story: name and switch custom themes through /theme, hand-edit JSON files in ~/.claude/themes/, and ship themes from plugins via a themes/ directory.
For the platform builders: hooks can now invoke MCP tools directly via type: "mcp_tool", which collapses a whole class of subprocess-shim hacks people had been writing to call MCPs from PreToolUse and PostToolUse. DISABLE_UPDATES environment variable now blocks every update path including manual claude update — the missing piece for IT-managed installs in regulated environments. WSL on Windows can inherit Windows-side managed settings via the new wslInheritsWindowsSettings policy key, which solves the dual-config-file problem that has plagued enterprise WSL rollouts. Bug-side: same-message-at-two-positions on fullscreen scroll (iTerm2, Ghostty, DEC 2026 terminals) is finally fixed, and the Write tool runs 60% faster on large files with tabs/&/$. claude upgrade.
Managed Agents First-Cohort Numbers Land — Notion, Rakuten, Asana on the Reference Page; $0.08/Session-Hour Holds
Two weeks after the April 8 public-beta launch, the Managed Agents reference page now lists Notion, Rakuten, and Asana as the named first cohort, plus a longer roster of unnamed enterprise users. Pricing has held at standard token rates plus an additional $0.08 per active session-hour for runtime — the “sandboxed code execution + checkpointing + scoped credentials + end-to-end tracing” bundle. The story for builders: this is the production stack you would otherwise be assembling out of Lambda, Fargate, Secrets Manager, and OTel. The bet Anthropic is making is that the integrated story is worth the per-hour premium for teams that don’t want to operate their own agent runtime.
Pair this with the ant CLI from yesterday and Routines from last week and the platform shape is coherent for the first time. ant is the developer surface (versioned prompts, tools, agent configs as YAML); Managed Agents is the runtime (sandbox, credentials, tracing); Routines is the trigger layer (cron, webhooks, repo events). For teams running multi-step agents in production today, the integration test is whether the three of them compose end-to-end without escape hatches — especially around debugging when a Routine-triggered Managed Agent does something unexpected. Worth a half-day this week if you have any agent infrastructure on the roadmap.
One Week to the April 30 Deadline — If Your CI Doesn’t Have a Model-Allowlist Yet, Today Is the Day
The 1M-token context beta on Sonnet 4.5 and Sonnet 4 expires next Thursday. Five working days. If you have not already done it: (1) grep your codebase for the context-1m-2025-08-07 beta header and either move those paths to Opus 4.7 (1M-context as a first-class capability, no header) or scope the prompt to fit 200K; (2) replace any unversioned claude-sonnet-4 or claude-sonnet-4-5 pins with dated versions so you control your own migration window; (3) add the model-allowlist to CI so future sunsets surface as test failures instead of 2am pages.
Also worth noting given today’s NEC announcement: enterprise rollouts at this scale are going to drive faster pin-disciplining across the ecosystem, because the customer-side procurement requirements (audit logs, version-pinning, change-control) come along for the ride. If you supply software to any of these accounts, your model-management posture is about to get reviewed even if you’re not the one building the agent. Opus 4.7 is the default landing pad and the migration guide is current.
GIC and Anthropic Co-Host Asia Pacific Innovation Day in Singapore — 150 Senior Leaders, Closed-Door Demos, and a 5.5x Per-Capita Number
GIC and Anthropic co-hosted an Asia Pacific Innovation Day in Singapore this week, bringing roughly 150 senior tech and investment leaders for closed-door demonstrations of Claude across enterprise scenarios. GIC senior vice president Dominic Soon used the keynote to publicly disclose that Singapore’s per-capita Claude usage is running at 5.5x the rate that population would predict — a number GIC clearly believes argues for Singapore as the regional Claude hub. GIC led the $30B Series G in Anthropic earlier this year and is now actively shaping the regional adoption playbook.
Combine this with today’s NEC announcement and you can see the shape of Anthropic’s 2026 APAC plan. Singapore is the financial-services and government-tech beachhead through GIC’s relationships. Japan is the enterprise-engineering beachhead through NEC. Korea (Seoul office, third in APAC) and Australia (Sydney office, fourth) are filling in the rest. The notable absence in this picture is China — Anthropic is not playing there and the Singapore numbers may be partially explained by regional firms that would otherwise be Beijing or Shanghai-centered routing through SG. The strategic question is how long the China-absent positioning holds before a competitor builds an equivalent hub from a different starting point.
The CISA-vs-NSA Mythos Split Hardens — America’s Cyber Defense Agency Is Still Locked Out
Axios and Cybernews confirmed through Tuesday and Wednesday: the Cybersecurity and Infrastructure Security Agency — the U.S. government’s lead cyber-defense organization for civilian critical infrastructure — does not currently have access to Mythos. The NSA does. The intelligence community has parts. CISA does not. Multiple sources put the explanation on the same Pentagon supply-chain risk classification that Trump suggested Monday was now being reconsidered. The friction inside the U.S. government is now visible across two news cycles and is starting to embarrass the administration.
The optics matter for the upcoming Pentagon negotiation. Anthropic refused to bend on the autonomous-weapons and domestic-surveillance lines in September, the Pentagon classified them a supply-chain risk in early March, the NSA started using Mythos anyway, CISA was left out, and now the White House is back at the table. The conventional read is that the deal closes with carve-outs that look much like Anthropic’s September position. The non-conventional read is that the carve-outs are written narrowly enough that Mythos-class capability flows to NSA and parts of the IC but not to civilian-defense CISA — which would be the worst-case outcome for the “defenders inside the tent” story Anthropic has been telling.
Last Night’s MCP Apps Outage on Claude.ai Was Resolved at 02:09 UTC — Six-Hour Window, No Data Impact Reported
Per the Claude status page: MCP apps were unavailable on Claude.ai for a window late Wednesday into early Thursday morning, with the all-clear posted at 02:09 UTC. The disruption affected MCP-served tool calls inside Claude.ai sessions; the API surface and Claude Code MCP servers were not implicated. No customer data impact has been reported. For teams that hit the window during late-evening U.S. or early-morning EU work, this is the explanation if you saw your connectors throwing connection errors before midnight Eastern.
The smaller signal worth recording: this is the second MCP-side incident on Claude.ai in the last fourteen days, after the connector-discovery flap on April 11. The pattern, if it holds, points to capacity or routing issues on the Claude.ai-side MCP serving stack rather than anything model-side. Not a crisis on its own, but if you operate in environments where you need a stable MCP endpoint for non-Claude.ai consumers, the API and code.claude.com surfaces have been more reliable.
Mythos Is Quietly Becoming Critical Infrastructure — And Nobody Has the Right Vendor Model for That Yet
Three weeks ago Mythos was a 40-organization defender preview. Today it is being shipped to G7 systemically-important banks within days. Australia’s federal government has a direct working relationship. The Reserve Banks of Australia and New Zealand are coordinating with global regulators. Japan’s FSA pulled MUFG, SMFG, Mizuho, the Bank of Japan, and the Tokyo Stock Exchange into a single meeting last Friday. The Bank of England is hosting bank-level discussions. The German Bundesbank president is publicly arguing for level-playing-field universal access. This is what the early stages of designating something critical infrastructure look like, except nobody is calling it that and Anthropic’s vendor relationship with each counterparty is still — functionally — a sales contract.
The third-party vendor breach this week is the operational signal that the existing model is undercooked for what Mythos is becoming. A contractor login should not be the failure mode for a tool that the central bank of England wants in production at the four largest UK banks. The right shape of the next thirty days is probably some combination of: (1) a published vendor-management standard from Anthropic with audit attestations, (2) bilateral information-sharing agreements with regulators in the major jurisdictions, and (3) some form of national-security-letter-equivalent process for handling contractor and vendor access logs. Anthropic gets to write the first draft of all three, but only if it moves before a regulator does. The NEC announcement today and the European bank rollout this week are reminders that the timeline is not abstract — this is the week the “narrowly gated frontier model” framing visibly stops being accurate.