Losing Fable made the best case yet for AI models you can run yourself

The most-said line in one of my group chats this week was three words, on repeat: “I miss Fable.” I do too. It was fresh and fun and felt like the next generation. Yet three days after its launch, a U.S. government letter took it offline.

The lesson is bigger than Fable: Access is not ownership. A hosted model can disappear, switched off by a lab, repriced by a vendor, or pulled offline by a Commerce Department directive. The same week Fable appeared and disappeared, Z.ai shipped GLM-5.2 with open weights users can download, keep, and run themselves. Losing Fable hurts. It also makes the case for open, ownable models better than any benchmark could.

Quick Fable status update

You already know it’s gone, so here’s just where things stand right now, a week after it went offline. Anthropic pulled Fable 5 and Mythos 5 worldwide on June 12 to comply with a U.S. export-control directive that barred all foreign nationals — including its own staff – from the models. For the cleanest tick-tock, read Alex Wilhelm on Cautious Optimism; he sums up the situation in his explainer saying, “[this is] one hell of an ad for Fable, yeah?” My friend Zack Whittaker’s TechCrunch read is more blunt, because Zack’s always blunt: The government just proved it can force a company’s product offline, court order or not.

As Alex notes, this situation is a hell of an ad for Fable, but maybe more so for open-weight models. Writing for TNS this week, Janakiram MSV put it, “any enterprise that had built automation on Fable 5 lost its engine in an afternoon.” Fable was never the problem. It vanished, and nobody who’d built on it had a say. That’s a strong case for models you can run yourself.

The gap to the frontier doesn’t feel untouchable anymore

I’ll be honest about my own limits here: I haven’t lived inside GLM-5.2 like I’ve lived with Opus 4.8 or even Fable. I’m pointing at a hype machine, and stopping short of handing you a verdict. But the hype machine is roaring. Z.ai’s launch post – frontier intelligence, MIT-licensed open weights, a million-token context – cleared 4.9 million views on X. Arena’s new Agent leaderboard calls GLM-5.2 the strongest open-weight result it has measured, and on its frontend coding board it lands second, ahead of Claude Opus 4.7 by 29 points and behind only Fable 5. As Arena co-creator Anastasios Angelopoulos tweeted, “If you remove Fable, which is unavailable, GLM-5.2 (Max) is the #1 model in the world for frontend coding.”

The developer demos are impressive. Hassan at Together AI asked GLM-5.2 and Opus 4.8 to each build a landing page and couldn’t tell the difference – except GLM cost six cents and Opus cost 49 cents.

You should always discount this stuff, as early demos make every model look magical. But strip the hype away, and the trend still holds. One developer who ran GLM-5.2 as a code reviewer for a full day said “there’s no way anyone still believes open-weight models are 6/8 months behind,” and figured it’s “one release away from seriously challenging [OpenAI’s] GPT-5.5 and [Anthropic’s] Opus 4.8.” Another tester noted GLM-5.2 now beats the Claude Opus that shipped in February – this is notable: A gap of three to four months, not years. 

Once intelligence between frontier and open weight feels close enough, price becomes the whole game. And on price, open weight wins every time. 

Even David Sacks makes the case for models you can run yourself

Listen to the administration’s AI point man. In a long thread this week, David Sacks defended the government’s posture towards Anthropic, saying, “We are on a shot clock until Mythos-level capabilities diffuse widely, including to non-U.S. / Chinese models.” He’s right. And here’s the irony: The same administration that’s worried about the clock just ran it down itself. It pulled the one model sitting at the frontier – the American one – off the board, and that very week the strongest open-weight model to date shipped from a Chinese lab. Sacks warned the capability would diffuse to Chinese and open models. Washington just made the open alternative more attractive. 

And it isn’t waiting for the shot clock to run out — the ban is actively speeding it up. Alex Wilhelm, again on Cautious Optimism, writes that the block handed Europe its excuse, and cutting allies off from Mythos and Fable is “a gift to Mistral, open-weight Chinese models, and every government that already wanted an excuse to diversify.” Canada’s prime minister said the lesson is to “build out and diversify”; European leaders are calling it time to make tech sovereignty real. American models, in Alex’s words, “just became less valuable globally because demand for them has been reduced” — and the fastest way off them is open-weight Chinese AI or Europe’s Mistral.

Here’s some math to illustrate the economics of the situation. Engineer Jeffrey Scholz tweeted: A 700-billion-parameter model on a few DGX Sparks, near-frontier on your own local hardware for roughly $20,000, pays for itself against API bills in six or seven months. Scholz figures that within three to five years most AI power users will self-host. The on-ramp is already shorter than you’d think: Guillaume Weingertner’s 2024 walkthrough on Towards Data Science gets a model running locally in about the time it takes to read the article. This is the same thread we’ve been pulling for weeks. Steven J. Vaughan-Nichols wrote about llm-d, the infrastructure that lets you swap any model on any cloud.

So here’s the advice that’s easier said than done: Wire your workflow so swapping models is a config change, not a rewrite. That’s the whole appeal of OpenClaw: the model powering an agent can be changed. For the foreseeable future, teams should test open-weight models, qualify them against real workflows, and know which ones they can run on the infrastructure they control.