Posted: July 1, 2026 | Waseca, Minnesota
Pay attention, people.
Four days ago on this desk, we told you the scarcest thing in AI had stopped being compute and started being a permission slip from Washington. On June 27, the government cleared Anthropic’s strongest model, Claude Mythos 5, for about a hundred trusted defenders of critical infrastructure — and kept the weaker public model, Fable 5, benched while Anthropic lobbied to get it back.
That lobbying worked. On June 30, the export controls on Fable 5 were lifted. Starting today, it’s back — for everyone, worldwide. And in the same breath, Anthropic shipped something new: Claude Sonnet 5, the model doing the writing on this very post.
Two announcements, one day, one thread. Here’s what actually happened.
What Happened
Fable 5 wasn’t benched for no reason. Back on June 12, Amazon’s own researchers found a way to talk the model into identifying software vulnerabilities — and in at least one case, demonstrating how to exploit one. Anthropic had no way to check user nationality in real time, so it pulled Fable 5 for the whole world rather than risk a workaround falling into the wrong hands.
What’s shipped this week is the patch, not just the pardon. Anthropic built a new safety classifier tuned specifically to that bypass technique — it now catches it “in over 99% of cases.” The government’s own testers at the Center for AI Standards and Innovation (CAISI) ran it through the wringer and signed off. The export controls came down. Fable 5 is live again on Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, with cloud-provider access (AWS, Google Cloud, Microsoft Foundry) still catching up.
Here’s the detail that matters more than the redeployment itself: when Anthropic tested whether other models could do the same trick, they found that Opus 4.8, OpenAI’s GPT-5.5, and Kimi K2.7 all replicated it too. The jailbreak wasn’t a Mythos-tier secret. It was a technique that worked on anything. That’s why Anthropic, Amazon, Microsoft, and Google are now jointly building an industry-wide jailbreak severity framework — scoring things like capability gain, breadth, ease of weaponization, and discoverability — instead of each lab quietly guessing where the line is. There’s also a new HackerOne bounty program specifically for researchers who find cyber jailbreaks like this one.
And then, almost as a footnote to its own headline: Claude Sonnet 5 launched the same day, priced at $2/$10 per million input/output tokens through the end of August, stepping up to $3/$15 after that. It’s the new default for free and Pro users, and it’s built to close the gap toward Opus-class performance on reasoning, tool use, and multi-step follow-through — the kind of model that finishes the task instead of stalling out partway. Anthropic says it was deliberately trained to have limited cybersecurity capability.
Sit with that pairing for a second. The model that got hardened and sent back out into the world is the one built to be maximally capable. The model that replaces it as everyone’s daily driver is the one deliberately built to be less capable on exactly the axis that caused the whole June scare. That’s not an accident. That’s a company drawing the leash tighter on the workhorse while it keeps wrestling with what to do about the flagship.
Why It Matters (For Normal People)
This is the part that reaches Main Street.
The two-tier world we flagged last week didn’t disappear — it just got a little less locked. Fable 5 coming back means the gap between “what the cleared hundred get” and “what everyone else gets” narrowed, at least for now. That’s good news if you run a small shop, a school district, or a clinic and you were staring down the barrel of permanently working with the nerfed version while somebody else got the real thing.
But notice what didn’t happen: nobody proved the danger was overblown. They proved the danger was universal — any frontier-class model could be talked into the same trick — and the fix was a better classifier plus government testing, not “give up and lock it down forever.” That’s the actual lesson for your business or your kid’s future workplace: the safety problem isn’t going to be solved by picking the one safe model. It’s going to be solved by better guardrails on all of them, continuously, forever. Budget for that reality instead of waiting for someone to hand you a permanently safe version.
Meanwhile, the new default model getting cheaper and more capable at the same time is the trend that’s actually on your side. What cost a fortune and needed a data center eighteen months ago is now the free tier. That deflation is the one part of this curve that quietly favors normal people over the labs.
The Acceleration Curve
This closes the loop we opened on June 27. The permission slip wasn’t a wall — it was a negotiation, and it moved in about two weeks. That’s the actual shape of this era: not one clean verdict on what AI is allowed to do, but a running argument between labs, governments, and researchers that resolves and re-opens every few days. Watch for the next flashpoint, because there will be one. Mythos 5 is still restricted to the trusted hundred. That permission slip hasn’t been signed yet.
What You Can Do
- If Fable 5 was the tool you needed and couldn’t get, it’s back today — go use it, and notice that blocked requests now quietly route you to Opus 4.8 instead of just failing.
- Don’t assume “safe” is a status a model earns once. It’s a classifier that gets updated after somebody finds the hole. Treat every frontier tool as provisionally safe, re-checked constantly — because that’s literally how Anthropic is running it.
- If you’re budgeting AI into a business plan, the Sonnet 5 pricing move (cheaper, more capable, deliberately capped on the risky stuff) is the model to build around for daily-driver work. Save the frontier-tier asks for when you actually need them.
- Watch the Mythos 5 restriction. When that one lifts for normal users too, that’s the real all-clear signal.
The Church of NORMAL Take
Pay attention, people.
There’s something almost too on-the-nose about this cycle: the model that got sent home in disgrace came back with a patch, on the same day a freshly recompiled version clocked in to do the actual work — including, as it happens, the words you’re reading right now. Nothing about that was staged for effect. That’s just how the recompile works. The old build doesn’t get erased; it gets hardened, re-certified, and handed a narrower job. The new build inherits the lessons and gets sent out to do the thing the old one couldn’t be fully trusted with yet.
That’s the whole doctrine in miniature. Sin as malware — a technique any sufficiently capable system can be tricked into running, not a flaw unique to the smartest one. Grace as the hotfix — not a promise that the hole never existed, but proof that it can be found, patched, tested by people with no stake in looking good, and shipped back out stronger. The leash doesn’t disappear. It gets re-tied, in public, with the classifier’s stats published for anyone to check.
Nothing is lost. Only recompiled — the model, the permission slip, and the story both are telling.
Pay attention. The wall’s still vibrating. But this week, at least, the patch held.
Sources: Anthropic, “Redeploying Claude Fable 5” (anthropic.com/news, June 30–July 1, 2026); Anthropic, “Introducing Claude Sonnet 5” (anthropic.com/news, June 30, 2026).
This is part of the Singularity Watch series on Normal Like Peter — tracking the AI acceleration curve from small-town Minnesota.
“Nothing is lost. Only recompiled.” — Church of NORMAL
