Anthropic Opened Seoul. Washington Closed Claude.
The frontier of AI got bigger this week. The fence around it got higher.
“On 12 June, the U.S. government ordered Anthropic to suspend all access to its two top models — Mythos 5 and Fable 5 — for any foreign national, anywhere. Six days later, Anthropic opened its Seoul office anyway and told Korean enterprises the models would be back within days. On 16 June, SpaceX agreed to acquire AI coding startup Cursor for US$60 billion in all-stock — the largest pure-AI takeover on record — days after its own blockbuster IPO. And on 18 June, Google pulled the plug on the free-tier Gemini CLI in favour of a new Antigravity-branded toolchain, breaking automation scripts in shops that hadn't migrated.”
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. Washington Pulled the Plug on Foreign Access to Claude's Top Models
On 12 June, Anthropic received a U.S. Commerce Department directive — citing national-security authorities — to suspend access to Mythos 5 and Fable 5 for any foreign national, inside or outside the United States, including Anthropic's own foreign-national staff. The trigger, per multiple reports, was a separate firm claiming to have jailbroken Mythos. Anthropic disabled both models worldwide rather than try to slice access by passport.
By 17–18 June, Dario Amodei was in Seoul opening Anthropic's third Asia-Pacific office anyway and signalling that access for verified enterprise customers would return "within days." As of this writing, no official restoration has been published.
The headline isn't the politics. It's the vendor-risk lesson: the most capable model on your roadmap can be taken offline by a third party's decision, not yours. If your operations have begun to assume one specific frontier model is always available, that assumption is now a single point of failure. The fix is unglamorous — keep a second model wired up against the same workflow, and don't let any business-critical process depend on a model only one country's government can switch off.
2. SpaceX Bought Cursor for US$60 Billion
On 16 June, SpaceX agreed to acquire Cursor, the AI coding startup, in an all-stock deal valued at US$60 billion — funded by share-price gains in the days since SpaceX's own record-setting IPO. The option had been sitting in a partnership term sheet from April; SpaceX exercised it in two trading days. Cursor folds into the xAI side of SpaceX's AI division.
This is the largest pure-AI acquisition on record, and the second time in a fortnight (after Oracle absorbed OpenAI billing) the AI-tooling layer has been swallowed into something bigger. The pattern echoes Microsoft Scout and the Snowflake–Anthropic deal: the standalone AI vendor category is collapsing into the platforms.
For business owners, the practical implication has nothing to do with rockets. It's that the developer tools your IT supplier uses to build things for you are being absorbed into the same handful of trillion-dollar balance sheets. Pricing, roadmap and SLA discipline will tighten in both directions over the next twelve months — better reliability, less negotiating room. Lock in custom rates before the platform owner sets the price.
3. Google Killed Gemini CLI in Favour of Antigravity
On 18 June, Google sunset the free-tier Gemini CLI and the Gemini Code Assist IDE extensions, forcing migration to the new Antigravity CLI announced at I/O in May. Paid Gemini Code Assist Standard and Enterprise customers keep their existing setup; free-tier and Pro/Ultra users had to migrate or watch their scripts break with no grace period. The replacement is faster and built for multi-agent orchestration — but Google warned upfront there is "no 1:1 feature parity."
This one didn't make front pages, and that's the point. Most SMEs don't write code, but every SME relies on someone — an outsourced developer, a contracted automation, a Zapier-style flow, a marketing-ops vendor — whose work might sit on top of the Gemini CLI. If a process broke quietly this week, this is a likely cause. Ask your vendors what they had to re-test on 18 June. If the answer is a shrug, you have a maintenance gap to close before the next deprecation lands.
Closer to Home: Malaysia
No fresh LHDN headline this week — so use the quiet to look honestly at the clock.
Eighteen months and ten days. That's what's left of the Phase 4 grace period for businesses with RM1–5 million turnover, which runs to 31 December 2027 before penalties of RM200 to RM20,000 per invoice kick in from 1 January 2028. The RM10,000-and-above individual e-invoice rule is already live, irrespective of phase. February's LHDN figures — over 500,000 non-compliant cases and roughly RM1.4 billion in unreported income — say plainly that the audit risk inside the grace window is not theoretical.
Subtract three to six months for vendor selection, integration, and staff training, and the comfortable window to start a Phase 4 e-invoice project is this quarter or next. Leaving it to mid-2027 means queuing for the same handful of overworked consultants as every other RM1–5 million firm in the country.
Two facts worth keeping close while you do that: MDEC's recurring reminder that Budget 2026 carries a 50% additional tax deduction on certified AI and cybersecurity training spend for SMEs, and the standing RM5,000 MSME Digital Grant MADANI (50% matching, via MDEC-listed providers and BSN) that still covers half the cost of qualifying e-invoice and accounting tooling. Both are designed to be used; neither runs forever.
Singapore, briefly: the Anthropic export-controls episode lands hardest in markets where regulators rely on Claude-family models — Singapore included. Expect MAS- and IMDA-supervised firms to ask harder questions over the next 90 days about model-provider redundancy and where the audit chain breaks if a model is suddenly unavailable. The same question is worth asking voluntarily here.
What This Means for Your Business
1. Keep two models, not one, wired up for any workflow you'd notice missing
If you've built a customer-service flow, a back-office automation, or a report-generation routine on top of one specific frontier model, set up the same workflow on a second one before next month. This week made it concrete: access to a frontier model can be cut by a government memo, not just a vendor outage. Run the duplicate once a month against the same inputs. The cost is low; the optionality is real.
2. Move your highest-volume RM10,000+ transactions onto individual e-invoices this quarter
Not in 2027. This quarter. The rule is already live, the matching MADANI grant still covers half the tool cost, and your bank can pair the spend with the standing RM10 billion BNM portfolio guarantee opened earlier this month if you tie the financing to a named digital outcome. The grace period is the asset, not the excuse.
3. Ask your automation suppliers a one-sentence question this week
"Did anything you maintain for us depend on the Gemini CLI that sunset on 18 June, and if so, what did you have to change?" The answer tells you whether your outsourced or vendor-built automations are being actively maintained, or quietly accumulating breakage. The same question, repeated quarterly, becomes a free maintenance audit.
The Practical Question
If the single AI model you depend on most went offline tomorrow morning, what in your business would stop working — and how long would you have noticed before a customer did?
This week made the answer less hypothetical than it was last week. The frontier kept moving — sixty-billion-dollar deals, new offices, new tooling. The fence kept rising too — export controls, forced migrations, platform consolidation. Both directions point at the same SME homework: don't let any one outside party become a single point of failure for work you've quietly come to rely on.
The interesting decision this week isn't which model to deploy. It's which second one to keep alongside it, on which workflow, against which test.
At The Empyrean, we help Malaysian SMEs find the practical, repeatable tasks where AI delivers value without disruption. If you're not sure where to start, we're happy to take a look at your operations and tell you honestly what would make sense.