This Week in AI: June 12–18, 2026

Six days. That's how long two of Anthropic's most capable models have been offline because of a single letter from the Commerce Department. In the same six days, SpaceX bought a code editor for $60 billion, OpenAI's real financials leaked and they're worse than anyone guessed, and GitHub got so overloaded by AI agents that Microsoft had to borrow servers from Amazon to keep it online.

Buckle up. Here's the week.

The 60-second version

• 🔌 A government order took Claude Fable 5 and Mythos 5 offline on June 12 — and they're still down, six days later.
• 💰 SpaceX bought Cursor for $60 billion, four days after its own record IPO, instantly becoming the 4th most valuable company in the US.
• 📉 OpenAI's leaked 2025 financials: $13B in revenue, $34B in costs, a $38.5B net loss.
• 🛠️ GitHub nearly fell over. Microsoft is now routing GitHub traffic through AWS — its biggest cloud rival — because AI agents are generating more load than Azure can absorb.
• ⏰ Gemini CLI dies today, June 18. If your scripts call gemini, go check them before your CI breaks.

Let's get into it.

A government order took two Claude models offline, and the fight over why is just starting

On June 12, at 5:21 PM Eastern, Anthropic received a letter from Commerce Secretary Howard Lutnick. It cited national security authorities and ordered an immediate halt to all access — by any foreign national, anywhere, including Anthropic's own foreign-born employees — to Claude Fable 5 and Claude Mythos 5. Both models had launched just three days earlier.

Anthropic couldn't selectively block by nationality in real time, so it shut both models down for everyone. Every other Claude model, including Opus 4.8, stayed live.

What actually triggered this? Per Anthropic's own statement, the government believes it learned of a way to jailbreak Fable 5 — a technique that could unlock some of the cybersecurity capabilities baked into the underlying Mythos model. Anthropic pushed back hard on the framing: it says the jailbreak is narrow, not universal, ties to a handful of already-known, minor vulnerabilities, and that other public models — it specifically named OpenAI's GPT-5.5 — are exposed to the exact same issue without facing any export controls at all.

"We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people." — Anthropic

Fortune reported that the order arrived the same week Anthropic confidentially filed for an IPO off a $965 billion valuation — timing that isn't doing the company's public-listing narrative any favors. It's also not the administration's first run-in with Anthropic this year: the Pentagon labeled the company a "supply chain risk" back in March, a designation Anthropic is fighting in court.

As of today, Fable 5 and Mythos 5 remain dark with no restoration date. If you had anything in production calling claude-fable-5, you're getting hard errors right now — the fix is swapping to claude-opus-4-7 or claude-opus-4-8, which keeps most code running with a step down in raw capability.

The lesson that actually matters here isn't political. It's that a frontier model can vanish from your stack with zero warning, for reasons that have nothing to do with your usage. If switching providers means rewriting your prompts from scratch, fix that problem this week, while it's still hypothetical.

SpaceX just bought Cursor. For $60 billion. In stock.

On June 16, SpaceX filed an 8-K with the SEC confirming an all-stock merger to acquire Anysphere — the company behind Cursor — at an implied equity value of $60 billion. Bloomberg reported the stock jumped roughly 17% on the news, vaulting SpaceX past Microsoft and Amazon to become the fourth most valuable company in America. Closing is targeted for Q3 2026, pending regulatory approval.

This isn't a spontaneous move. Reuters reported SpaceX had an option from an April partnership to either acquire Cursor outright for $60 billion or pay a $10 billion fee to keep working together. It chose to buy.

The number that justifies the price tag: Cursor's annualized revenue went from $1 billion in November 2025 to roughly $4 billion by June 2026 — 4x in about seven months. At 15x ARR, $60 billion stops looking reckless and starts looking like a land grab before the price goes up further.

It also closes out a consolidation cycle that's moved fast. Every major AI coding tool is now owned by one of five companies:

Tabnine is the only serious AI-native coding tool left standing without a giant behind it. If your workflow leans on an "independent" AI dev tool right now, it's worth asking who owns it in twelve months.

OpenAI's real 2025 numbers leaked, and they're rougher than the growth story suggested

Journalist Ed Zitron published audited financial documents he says were independently verified by the Financial Times, and the numbers are stark: OpenAI posted $13.07 billion in revenue against $34 billion in total costs for 2025, landing on a net loss attributable to the company of $38.5 billion — about 7.5x its 2024 loss.

Strip out a one-time $41.55 billion non-cash charge tied to OpenAI's nonprofit-to-for-profit conversion, and the underlying operating loss was $20.92 billion — still a company spending $1.60 for every dollar it earns. R&D alone hit $19.18 billion, more than the company's entire revenue. And Microsoft pocketed $17.2 billion of that spend — making Microsoft simultaneously OpenAI's biggest investor, its cloud provider, and its single largest expense line.

This drops eight days after OpenAI confidentially filed its draft S-1, in the same general window where Anthropic is reportedly preparing its own filing fresh off a $965 billion valuation. Goldman Sachs projects 2026 AI IPO proceeds could hit $160 billion — a 4x jump from 2025 — driven almost entirely by SpaceX, OpenAI, and Anthropic. Three near-trillion-dollar AI offerings inside one calendar quarter has never happened before. Watch how each one prices relative to the others; that tells you more about where "AI bubble" talk actually stands than any new benchmark will.

GitHub nearly broke. Microsoft called in its biggest rival to save it.

Here's an uncomfortable one for Microsoft: Business Insider reported, and Microsoft confirmed, that GitHub has been routing traffic through Amazon Web Services because AI coding agents pushed the platform past what Azure could handle.

The growth curve explains why. GitHub's COO confirmed the platform is processing 275 million commits a week, putting 2026 on pace for roughly 14 billion commits total — up from 1 billion in all of 2025. AI-agent-opened pull requests jumped from 4 million in September 2025 to over 17 million by March 2026. GitHub Actions usage hit 2.1 billion compute minutes in a single week, up from 500 million weekly back in 2023.

The cost showed up as reliability problems: nine service-degrading incidents in May alone, and June availability sitting around 88.4% — well under the 99.9% enterprise SLA most teams assume is guaranteed. Microsoft's own infrastructure team had planned for a 10x capacity expansion; by February they realized they needed 30x.

Microsoft calls the AWS arrangement temporary while it accelerates GitHub's full migration to Azure, targeted for 2027. Whether that holds is a separate question — but if your CI runs have felt slower or flakier this month, this is why. It's not your YAML.

Gemini CLI dies today — here's the fast version

If anything in your stack shells out to gemini, this is the deadline: Google's own migration post confirms that as of today, June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions stop serving requests for Google AI Pro, Ultra, and free-tier users. The replacement is Antigravity CLI — rebuilt in Go, with async multi-agent workflows, sharing the same backend as the new Antigravity 2.0 desktop app.

Worth knowing before you panic-migrate:

• Enterprise is unaffected. If your org runs Gemini CLI through a Gemini Code Assist Standard or Enterprise license, or through Google Cloud, nothing changes today.
• No 1:1 feature parity at launch. Google says so directly. Agent Skills, Hooks, Subagents, and Extensions (now "Antigravity plugins") carry over; everything else needs testing.
• It's a fresh install, not an upgrade. Antigravity CLI is a separate binary, not a drop-in replacement for the old one.

If you've got gemini calls baked into cron jobs, Git hooks, or CI configs on a free or Pro/Ultra account, grep for them today — they fail silently with no warning banner, the command just stops returning anything.

Why everyone's borrowing everyone else's compute

Microsoft leaning on AWS isn't a one-off. It's a symptom. The same week, reporting surfaced that Google is paying SpaceX roughly $920 million a month for bridge AI compute capacity through 2029 — despite running one of the world's largest cloud businesses itself.

Every one of these companies competes directly with at least one of the others, and they're all now renting capacity from each other because nobody can build fast enough alone. Data centers take 18 to 36 months to come online. Demand isn't waiting around for that math to catch up.

What this means for you, the developer

• Don't single-source your model layer. Fable 5 went dark with zero notice. If a provider swap means a prompt rewrite from scratch, that's a design flaw you can fix now — cheaply — before it becomes an outage you can't.
• Audit anything that calls an external CLI today, not next sprint. Between the Gemini CLI sunset and the rolling API deprecations Google's pushing through June, something in a pipeline somewhere is going to fail quietly this week.
• The "independent AI tool" era is over. Five companies now own essentially every serious AI coding tool on the market. Build your stack assuming further consolidation, not assuming a small player's roadmap and pricing hold steady.

That's the week. See you next Friday.

Originally published on ZyVOP

💡 For more articles like this, subscribe to the ZyVOP newsletter!