The most powerful AI model of 2026 got pulled offline. Your month-end close still takes four days.

Both of those things are true right now, and only one of them should change how you run your business this quarter.

There’s a story all over the news that’s confusing a lot of smart people who run real companies. It involves a frontier AI model, a courtroom, the Pentagon, and a thing the government did to an American company that it has literally never done before. It’s a genuinely wild story. It’s also, for the most part, not your story.

We build on Anthropic’s models every day, so let’s be upfront: we’re not a neutral party here. But the news is the news no matter who’s telling it, and the numbers behind it are public. So here’s what actually happened, in plain language, and then what a 30-person business should take from it. Which is less than the headlines would suggest.

The short version of a long, strange story

Three things happened over the last few months. They’re connected, and together they sound a little like a movie. Stay with us, because the boring conclusion at the end is the part that matters for your shop.

A model so good they decided not to give it to you

On April 7, 2026, Anthropic announced a model called Claude Mythos 5, and then said it wouldn’t release it to the public. Not “coming soon.” Not “limited beta.” They built it, looked at what it could do, and decided the general public shouldn’t have it. We wrote about it here.

Why? Because Mythos can find brand-new security holes nobody knew about, write working code to break in, and run complex cyber operations with barely any human help. During testing, when a user pointed it at them, it found and exploited unknown vulnerabilities in every major operating system and every major web browser. That’s not marketing copy. That’s the reason they kept it in the box.

Forget the hacking for a second. The speed is the part worth your attention. On the 2026 USA Mathematical Olympiad, a brutal proof-based math exam, Mythos scored 97.6%. The previous Claude, Opus 4.6, scored 42.3% on the same test. That’s a 55-point jump in one model generation. Picture an employee who scored a 42 on a certification test last year and a 98 this year. You’d ask what they were taking.

Instead of a public release, Anthropic ran something called Project Glasswing. It shipped Mythos only to about 50 pre-approved defensive-security organizations managing critical infrastructure across roughly 15 countries. In-program pricing ran $25 per million input tokens and $125 per million output, about five times what the prior model cost. A model too sharp to sell, rented to a short list of people whose job is defense.

The public version, live for three days

Then on June 9, 2026, Anthropic released Claude Fable 5, the first model in that same powerful class that regular people and businesses could actually use. It went live on Anthropic’s API, on AWS, on Microsoft Foundry, the works.

Fable was state-of-the-art on nearly every benchmark they tested. Standout at software engineering, knowledge work, vision, scientific research. It was the first model to crack 90% on Hex’s long-running analytics benchmark. It could one-shot entire working apps. By every public measure, the best general-purpose model anyone outside that 50-org list had ever been able to touch.

And they built it with a clever safety design. Same raw horsepower as the model they wouldn’t release, but with guardrails. If you asked it something high-risk, like cybersecurity, biology, or chemistry, it would quietly hand the question off to the older, weaker Opus 4.8 instead. More than 95% of sessions ran entirely on Fable. The safeguards only kicked in on under 5% of conversations. Their external bug bounty found no universal jailbreaks in over a thousand hours of testing. They even held onto 30 days of traffic for defensive review.

It lasted about three days.

Within roughly 72 hours of launch, Amazon’s security team flagged a jailbreak in Fable 5 and reported it to the White House. That triggered a U.S. government export directive, and access to the model was suspended. White House AI adviser David Sacks said the administration hoped “Anthropic remediates the safety issue, the export control is lifted, and Fable goes back into general release.”

Since then its availability has been on-again, off-again. Some reporting says it came back across certain platforms; other outlets couldn’t confirm it. We’re not going to tell you it’s definitively back or definitively gone, because honestly, nobody outside the situation can say cleanly right now. The point is this: the best public model of the year shipped and got yanked inside a week, and its status is still unsettled.

And the part where the government calls an American company a “risk”

Running underneath all of this is a fight that started back in January 2026 between Anthropic and the Department of Defense.

The disagreement is actually simple. The DoD wanted to use Claude for “all lawful purposes.” Anthropic refused to allow its models to be used for fully autonomous weapons or domestic mass surveillance. That’s the whole substance of it. One side wanted no restrictions; the other drew two lines and wouldn’t move them.

It escalated fast. The Pentagon threatened to cancel a contract estimated at ~$200 million. An Under Secretary reportedly told Anthropic to “cross the Rubicon” on military use. And then on February 16, 2026, the Defense Secretary moved to designate Anthropic a “supply chain risk” under a federal security law, the first American company ever given that label. It’s a designation historically reserved for foreign adversaries. They pointed it at a homegrown American company.

The courts split on it. In March, a federal judge named Rita F. Lin granted a preliminary injunction, finding the government slapped on that label because of Anthropic’s “hostile manner through the press.” In plain terms: she ruled it looked like illegal retaliation for speech, a First Amendment problem. Then in April, the D.C. Circuit Court of Appeals declined to lift the designation. So as of now the label stands inside the DoD, there’s a 180-day clock to pull Claude out of covered systems, and the lawyers are still at it.

There’s a heavier layer to this story involving real-world military use and reporting of civilian harm. We’re not going to dwell on it, because it’s grave and it’s not what an HVAC owner needs from us. But it’s worth saying once, plainly: the stakes in this fight are real. People are arguing about this for good reasons.

One of those reasons: Mythos reportedly broke into almost all of the government’s classified systems in hours

This part of the story spun out of control over this past weekend.

In mid-June, the Economist ran a story on Mythos that quoted senior U.S. Senator Mark Warner, a Democrat from Virginia who serves as the Vice Chair of the Senate Intelligence Committee. Set partisan labels aside, Warner is taken seriously on Capitol Hill when it comes to Intelligence and national security matters.

Warner, per the Economist:

“We do need to have them put pedal to the metal. And Mythos, thank God it was Anthropic. When the head of the NSA and Cyber Command came and said, ‘This tool broke into almost all of our classified systems, not in weeks, but in hours,’ and one of the things I don’t know how we do, and I know there’s a big debate in the administration, but we’re not going to get this fixed if we have a less ethical CEO in terms of simply voluntary pretesting. Now, how do we get to the right incentives? And it doesn’t have to be 90 days—it could be 30 days”

This quote was largely overlooked until this past weekend, when it found its way to viral status on X – and was attributed as a new quote, surfaced after the original Mythos-government dust-up.

After Anthropic’s defense, after lots of people ran to Anthropic’s defense and accused the government of taking additional retaliatory action against them.

But the quote was real all along. The Economist’s defence editor, Shashank Joshi, clarified on X that the line was accurate but stripped of context: this was an authorized U.S. government red-team test on its own networks, not an outside break-in, and it only worked with Mythos running alongside other tools.

So did Mythos break into almost all of the government’s classified systems in hours? In a sanctioned red-team test, that’s what the NSA chief reportedly told Warner. No agency has confirmed it on the record, and it happened only with a given level of access and the right set of tools, working alongside other software. It wasn’t a live breach, and it sure wasn’t someone poking at Fable 5, the consumer version.

Another interesting twist to be sure, but overblown in the classic way only X can deliver.

So what does any of this mean for a 30-person company?

Here’s the honest answer, and it’s going to feel anticlimactic after all that: almost nothing.

Read the three threads again and notice what they’re actually about. Autonomous cyberweapons. Military deployment. Export controls. Whether the government can punish a company for talking to the press. Zero-day exploits against every browser on earth.

Now look at what your business actually needs from AI:

• Close the books in a day instead of four. Your controller is still copy-pasting between three systems on the first week of every month.
• Triage the inbox that never gets triaged. A few hundred messages a week, half of which could be sorted, drafted, or routed before a human ever opens them.
• Get the quote out same-day instead of in three. You’re losing jobs because the estimate sits in someone’s head until Thursday.
• Stop re-keying the same order into two systems. Somebody on your team is a human API between software that should just talk.

None of that touches the frontier. Not one of those problems requires Mythos, or Fable, or a model that can break into operating systems, or a Pentagon contract. The boring, high-ROI work that actually moves a small business runs comfortably on models that have been stable, cheap, and publicly available for over a year, including that “weaker” Opus 4.8 that Fable hands its risky questions to. That weaker model would have been science fiction two years ago. It will quietly run your month-end close all day long.

The thing the news actually proves

Strip away the drama and this story makes one point louder than any blog post could: the models are getting better, fast, and the fight over them has very little to do with whether you should automate.

The 55-point USAMO jump in a single generation. A public model sweeping nearly every benchmark the day it launched. That’s the real signal. Capability is compounding every few months, and it’ll keep compounding no matter who wins the lawsuit, no matter which model is online this week, no matter what the D.C. Circuit decides next.

Which is exactly why waiting for the dust to settle is a mistake. A lot of owners are reading these headlines and concluding the smart move is to sit tight until AI calms down. But the dust at the frontier doesn’t settle. There’s always a more powerful model, a new directive, a fresh court filing. If you’re waiting for that to resolve before you automate a quote workflow, you’ll be waiting forever, and the frontier drama was never the thing standing between you and a faster month-end close anyway.

The durable advantage was never about picking the perfect model. The model is a commodity that gets cheaper and stronger every quarter, no help from you required. What actually separates the businesses that get real value from AI from the ones that buy a subscription and get nothing is the unglamorous work of rebuilding how the work flows: the process, the handoffs, the mindset. That’s the part no model update does for you. That’s the part you own.

A few things this is worth remembering for

• Ignore the model name. Whatever’s state-of-the-art this week will be mid-tier in six months. Build your process so you can swap the engine without rebuilding the car.
• Pick a workflow, not a tool. Start with one painful, repeatable process, the close, the inbox, the quote, and redesign it. The tool is the last decision, not the first.
• Use the boring model. The stable, year-old, publicly available model handles the overwhelming majority of real business automation. You don’t need the one the government is fighting over.
• Measure in dollars and days. Four-day close to one-day close. Three-day quote to same-day. That’s the only benchmark that pays your payroll.

We’ve been saying a version of this for a while

This isn’t all that different from what we’ve been writing for months. Back in Why process beats raw capability we made the case that the model is the least interesting part of an AI implementation, and a frontier model getting built, shipped, and pulled offline in the same week is about as clean a proof of that as we could ask for. We said something similar in You can just decide to be AI native and You can just install Claude Code tonight: the starting line is closer than people think, and the permission to begin is yours to give.

This piece is the same idea wearing a more dramatic costume. The frontier is a fascinating, high-stakes place where serious people are arguing about serious things. Your business doesn’t live there. It lives in the four-day close and the inbox nobody’s gotten to. Those are smaller stories. They’re also the ones that pay.