Read time: 4 minutes

Welcome back, leader!

Anthropic just proved a model can be "back" and still come with an asterisk.
This week: what actually changed with Fable and Sonnet, why OpenAI just did the same dance, a free AI model worth trying this week, and a habit that saves you from re-explaining yourself to Claude every time.

Quick recap - Export controls are government rules that restrict who can access a technology, normally used for things like military hardware.
On June 12th, in one of the first known cases of that rule being applied to a general AI model, the government gave Anthropic 90 minutes to pull Fable 5 offline worldwide.

The reason: researchers at Amazon reportedly got Fable to help spot a software vulnerability and, in one case, write code showing how to exploit it.
Anthropic had no fast way to check every user's nationality in real time, so it shut the model down for everyone rather than risk breaking the rule.

18 days later, on June 30th, the government lifted the restriction.
Fable 5 came back July 1st, everywhere you'd normally use Claude.

The catch: Anthropic built a new filter (a classifier) that specifically watches for the trick that got Fable banned, and says it catches it over 99% of the time.
If your request trips that filter, you get quietly switched to a weaker model, Opus 4.8, instead, with a note telling you it happened!

Bottom line: Fable works again, but don't be surprised if some requests get quietly downgraded to a different model without you asking. Give it a try as Fable 5 is only included in the subscription plans until July 7th.

The same week, Anthropic made Sonnet 5 the new default for Free and Pro plans (not Max, Team, or Enterprise, those still default elsewhere). It costs less to run than Opus 4.8, and performs nearly as well on everyday work.

Worth knowing: a cheaper price per use doesn't always mean a cheaper bill.
I have found that Sonnet 5 takes more steps to finish a task, so the total cost can land closer to the pricier models than the sticker price suggests.

Bottom line: use Sonnet 5 for your daily work. Save Fable for the genuinely hard or high-stakes stuff.

This isn't just an Anthropic story. Two weeks earlier, the government asked OpenAI to do the same thing with its new flagship, GPT-5.6 Sol, restricting it to a small group of trusted partners instead of releasing it to everyone. OpenAI said publicly it doesn't want this kind of gatekeeping to become the norm.

Whether you're on Claude or ChatGPT, the pattern is the same: the newest, most capable model doesn't ship to everyone on day one anymore.

And this connects to the rest of this issue. Two days after Fable went dark, a different company released a free AI model that doesn't answer to any single government or boardroom. That timing wasn't a coincidence…

Here's the plain-English version. Claude is "closed": you access it through a subscription, and Anthropic controls what it can do and who gets it, exactly what just played out with Fable.

GLM-5.2 is "open": the company that made it, Z.ai, published the actual model for free. Anyone can use it, run it, or build a product on it without paying Z.ai anything.

Why should a business reader care? Two reasons:

First, cost: GLM-5.2 is dramatically cheaper to use than Claude or ChatGPT for most everyday tasks.

Second, and more relevant given what just happened with Fable and GPT-5.6: it's not controlled by any single government or company's decision about who gets access. If the tool you rely on can be switched off with 90 minutes' notice, a backup that can't be switched off the same way is worth knowing about.

How to actually try it, no setup required:

Go to chat.z.ai in your browser and use it exactly like you'd use ChatGPT.
Type a question, get an answer. There's a free tier before it starts charging you, and even the paid tier runs a few dollars a month for normal use. That's the whole instruction. No installs, no code.

If you have a developer on your team who's curious, GLM-5.2 also plugs directly into the coding tools they already use, at a fraction of the cost of Claude or GPT. One developer built the same small app three times: once with Claude Code, once with GLM-5.2 through a free route, and once with GLM-5.2 through a paid route that still cost only 35 cents total. All three apps worked and looked nearly identical. Worth forwarding to whoever owns your dev tooling budget, even if that's not you.

Bottom line: open chat.z.ai this week and ask it something you'd normally ask Claude. Ten minutes, no commitment, and you'll actually understand what people mean by "open" AI instead of just nodding along.

Someone ran the same prompt through three AI models and compared what each one produced in a single try. The interesting part is that a smaller, compressed version of GLM-5.2 held up surprisingly well against the full-size models, on the first attempt, running entirely on a home computer instead of a data center!

If you keep re-explaining the same task to Claude every week, you're doing the boring part twice. Here's a way to stop.

Pick one thing you do on repeat: a variance summary, a client update, a weekly report.

Write down the messy version, your steps, your examples, the format you want, the things you always end up correcting.

Then paste this into Claude:

Save what it gives you. Next time, instead of retyping three paragraphs of context, you just say "use my [Skill Name] Skill on this input."

Bottom line: try this on one task this week before doing anything more ambitious with it.

A Skill is basically a saved instruction sheet Claude follows every time.
Once you have one working, you can chain a few together, or bundle them into something that runs on a schedule without you touching it. That's a project for another week.