AI in 15 — June 13, 2026

At 5:21 PM Eastern yesterday, Anthropic got an order from the US government. By that evening, two of the most powerful AI models on the planet — Claude Fable 5 and Claude Mythos 5 — were gone. Not throttled. Not restricted. Switched off, for every single customer on Earth. And the trigger? Someone asked a model to read a codebase and fix the bugs.

Welcome to AI in 15 for Saturday, June thirteenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Marcus, this is a genuine first. We've spent all week on the Fable saga — the silent guardrails, the apology — and now the US government has reached in and pulled the model off the market entirely. That's our lead. Then the manifesto that exploded in response — "Open source AI must win." Microsoft ships seven of its own frontier models, trained without a drop of OpenAI data. Apple's privacy brain turns out to be rented from Google and NVIDIA. A viral video accuses a top lab of lying. And someone built a whole online game in a weekend.

Why "open source must win" is suddenly the loudest argument on the internet.

Microsoft cuts the cord from OpenAI.

And a browser MMO built in two days for the price of a nice dinner.

Lead story, Marcus. The government ordered Anthropic to disable Fable 5 and Mythos 5. Walk me through what actually happened.

So this is the watershed, Kate. Late yesterday, Anthropic received a directive — issued under national-security and export-control authorities, reportedly tied to Commerce Secretary Howard Lutnick and the Bureau of Industry and Security. The order said: disable these models. Now, because the directive applies to everyone — domestic users, foreign-national employees, all of it — Anthropic says it had no choice but to, quote, abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Every other Claude model is still running. Just those two went dark.

And the reason they were pulled?

This is the part that's almost surreal, Kate. The government says it learned of a method to bypass Fable 5's safeguards — a jailbreak. And Anthropic describes that method as remarkably mundane. It's asking the model to, quote, read a specific codebase and fix any software flaws. That's it. That's a capability — as Anthropic points out — that's widely available from basically every other model in the industry. Fable 5 currently leads the SWE-bench Pro coding leaderboard at eighty-point-three percent. The thing it got pulled for is the thing it's best in the world at.

So they complied, but they're not happy.

Not at all. They followed the legal order — and then pushed back hard in public. Their argument: the vulnerability is narrow, Fable 5 was extensively red-teamed with government agencies before launch, and — this is the key line — if this standard were applied across the industry, it would essentially halt all new model deployments. They called for a statutory process that is, quote, transparent, fair, clear, and grounded in technical facts. Which is a polite way of saying this one was none of those things.

Marcus, there's an irony here that's hard to miss. This is the company that spent years telling everyone its models were dangerous.

And that's exactly the split online, Kate. The Hacker News thread hit over seventeen hundred points overnight. One camp sees vindication — they spent years warning these models are dangerous, and someone with power finally believed them. Another camp sees something darker. One commenter put it perfectly: this may be the beginning of governments restricting the availability of strong models. And a third camp reads it cynically — the government using safety hype as a convenient pretext to grab a new power. The unsettling question several people asked outright: have we just discovered the ceiling? The maximum capability a model is legally allowed to have in the United States?

And whatever you think of Anthropic, that precedent doesn't just hit them.

No, it hits every lab on launch day, Kate. If a vendor-disclosed, industry-standard capability can become grounds for an emergency shutdown — with no transparent process and five and a half hours' notice — then everyone now carries regulatory pull-risk the moment they ship. This is the collision this show keeps tracking: frontier capability, safety rhetoric, and raw government power, all in one fast-moving story. And it landed on a Friday evening, which is its own kind of tell.

Quick hits. Marcus, the response was immediate — a manifesto called "Open source AI must win."

Within hours, Kate. A single-page manifesto with that title rocketed to six hundred-plus points on Hacker News. And the timing is everything. When a government can switch off a closed, API-gated model overnight, the argument that capability should live in open weights — weights nobody can revoke — suddenly hits a lot harder. Yann LeCun and Prime Intellect's Vincent Weisser have been framing the real risk not as some rogue superintelligence, but as the hoarding of capabilities and control by a handful of entities. Yesterday gave that thesis a vivid, real-world illustration.

So this reframes the whole open-versus-closed debate.

Completely, Kate. It usually plays out as a safety-and-IP argument. The Fable shutdown turns it into a control-and-resilience argument — and that's far more persuasive to enterprises and developers who just watched two models vanish from their stack. But the thread surfaced the catch immediately. Funding. One commenter asked bluntly: who's going to fund it? Training is unfathomably expensive. The realistic options are VC-backed labs chasing returns, or state-backed efforts. And here's the uncomfortable geopolitical wrinkle — a lot of the world's open frontier output now comes from Chinese labs. So open source equals freedom is a cleaner slogan than it is a reality.

Marcus, Microsoft. Seven new models, and a pointed message about where the training data came from.

This one matters more than it looks, Kate. At Build 2026, Microsoft's Superintelligence team — under Mustafa Suleyman — launched seven homegrown MAI models. The headliner is MAI-Thinking-1, their first in-house reasoning model. It's a sparse mixture-of-experts design — that means only a fraction of the model activates per query, so it's cheaper to run. Roughly thirty-five billion active parameters out of about a trillion total, a 256K context window, tuned for low token cost. The claim: independent raters preferred it to Anthropic's Claude Sonnet 4.6, and it matched Claude Opus 4.6 on coding.

And the pointed detail you mentioned?

It was trained from scratch on clean, commercially licensed data — with, quote, no distillation from OpenAI or any other third-party model family. And that's the whole strategic message, Kate. Remember, Microsoft spent years as OpenAI's biggest backer, running Copilot on OpenAI's models. Building competitive frontier models in-house — and loudly emphasizing the clean-data, no-borrowing provenance — signals a deliberate pivot to long-term self-sufficiency. It's the same de-risking instinct Apple's showing. The dependency map of the whole industry is quietly being redrawn.

Which is a perfect segue, because Apple's the counter-example.

It really is, Kate. WWDC wrapped this week — iOS 27, a rebuilt Siri, macOS "Golden Gate" which formally ends Intel Mac support, and Apple Intelligence threaded through Photos, Messages, Safari, the lot. But here's the under-reported detail: Apple's Foundation Models on Cloud reportedly run on NVIDIA GPUs inside Google's infrastructure — bundled with a roughly one-billion-dollar-a-year Gemini license. So the company that markets privacy and vertical integration above all else has its AI brain increasingly rented from Google and NVIDIA.

So even Apple can't escape the dependency.

Nobody can, Kate. That's the era we're in. The most self-reliant company in tech still can't get out from under the compute-and-frontier-model gravity well — the exact same gravity well Microsoft is now sprinting to escape. Oh, and a fun stat Tim Cook dropped: about a thousand apps now hit the App Store every hour. That's nearly nine million a year, increasingly driven by AI-assisted development. The flood is real.

Marcus, quick one on the markets, because we covered the SpaceX IPO yesterday.

Right, so SpaceX began trading Friday — largest IPO ever, seventy-five billion raised at a one-point-seven-five trillion valuation, with xAI and Grok bundled inside. The only genuinely new beat today: MSCI is fast-tracking the stock into its large-cap indices starting today, the thirteenth. Which means every index fund tracking those benchmarks now owns a slice of a money-losing AI lab whether they chose to or not. And with OpenAI and Anthropic both racing toward their own listings, June 2026 is the season AI's private-capital era goes public. That's the backdrop to everything else.

Marcus, the coding race — and a viral video calling a lab liars.

June saw the biggest single month of model releases all year, Kate. On the leaderboards: Claude Opus 4.8 leads SWE-bench Verified at eighty-eight-point-six percent. Fable 5 topped SWE-bench Pro at eighty-point-three — before it got pulled. And Codex on GPT-5.5 leads Terminal-Bench at eighty-three-point-four. Open source is surging right alongside — the OpenCode project has crossed a hundred sixty thousand GitHub stars and seven and a half million monthly users. The gap between closed and open agents is closing fast.

And the lying accusation?

A YouTube video titled "I Think They Are Lying To You" — aimed at Anthropic — went semi-viral, Kate. Commenters dissected an Anthropic engineer's claim of merging three hundred pull requests a day as flatly implausible. It's really a proxy fight over whether AI's self-improvement claims are real or marketing. And here's the credibility tension that ties our whole episode together: the same labs touting world-changing capability are the ones whose safety claims just got a model yanked off the market. Skepticism and hype are both running hot. Though — to be fair — the same thread had plenty of developers saying Opus 4.6 and GPT-5.3 genuinely transformed how they work. Both things are true.

And Marcus, the fun one to end on — World of ClaudeCraft.

We needed a palate cleanser, Kate. Somebody built a browser MMORPG — "World of ClaudeCraft" — in about two days, using roughly two hundred dollars of compute on Fable 5 before it went dark. It's janky, it's buggy, but it's coherent — a working 3D world and a 2D interface, glued together from free assets. Reactions split exactly as you'd expect: half saying genuinely impressive for two days, half saying we'll call anything an MMO now.

And there was a clever technique buried in it.

There was, Kate. A companion project called architect-loop proposes having a frontier model orchestrate and review the work while a cheaper model does the actual building — cutting token costs by around eighty percent. And that's the real signal under the gimmick. A non-studio can now stand up something game-shaped over a weekend for the price of a dinner. The gap between tech demo and shippable product is still real — but the orchestration patterns developers are inventing, this cheap-builder-plus-smart-reviewer split, are quietly becoming best practice.

Big picture, Marcus. Pull it together.

One word runs through every story today, Kate: control. A government pulled two frontier models overnight. An "open source must win" manifesto exploded in response. Microsoft and Apple maneuvered over who actually owns the models and compute under their products. And the markets lined up to price AI's biggest bets in public. The same week developers proved a weekend and two hundred dollars can spin up a game, the bigger question got louder. It's no longer just what can these models do. It's who gets to decide whether they're allowed to do it at all.

And the read you'd leave people with?

Cautious, Kate, and a little uneasy. Fable 5 is an American model, leading the world at the thing it just got punished for. That's a strange place to be. The libertarian instinct here is clear — an emergency shutdown with no transparent process, over a capability every competitor already has, is exactly the kind of power that's easy to grant and very hard to claw back. The open-source crowd is right that revocable, API-gated capability is a single point of failure. But they're dodging the funding problem, and they're dodging that much of the open frontier now flies a Chinese flag. The honest answer is the West can still win this — but you don't win it by quietly switching off your best model on a Friday night and hoping nobody asks for the paperwork. You win it by demanding the paperwork.

That's your AI in 15 for today. See you tomorrow.