GLM-5.2 contra Claude Fable 5: què diuen les dades?

On June 9th, 2026, Anthropic released Cloud Fable 5, calling it the most capable model it had ever made public. Three days later, the US government shut it down. And by the following day, a Chinese AI lab had already created and released a model beating Fable 5 on some benchmarks. The model is GLM 5.2 from a company called Z.AI. The story here is the speed and precision with which Z.A.I. turned Anthropics regulatory crisis into proof of an argument it had been building for over a year. That frontier intelligence should be open and not licensed out by a government directive. Let's quickly look at what Anthropic did. Claude Fable 5 was Anthropics first methose class release. A tier the company said sat above

its existing opus models in raw capability. The system showed most strength in coding, scientific reasoning, and long-running autonomous tasks, the kind of capability that's useful and risky at the same time. Anthropic said as much itself, warning at launch that a model this capable needed serious guardrails, especially around cybersecurity. Three days after launch, the US Commerce Department invoked national security export control authority and ordered Anthropic to cut off Fable 5 and its underlying model, Mythos 5, for any foreign national, including Anthropics' own non-citizen employees. The directive didn't distinguish by location.

A foreign employee working inside Anthropics San Francisco office fell under the same restriction as someone logging in from somewhere else. The anthropic said it had no way to verify nationality in real time, so it disabled both models for every customer within hours of receiving the order. Anthropic disputed the reasoning. The company said the trigger was a narrow jailbreak, not a universal bypass of fables safeguards, and that the same class of vulnerability likely existed in other public frontier models facing no such order. That dispute is still unresolved, but what happened next didn't wait for it to be.

On June 13th the day after the directive landed, Z.A.I. announced GLM 5.2. Its fourth flagship release in roughly four months. The model runs on a mixture of experts architecture, 744 billion parameters in total, but only around 40 billion activate for any given task. The way to think about it is a large team of specialists where each query gets routed only to the relevant experts, rather than waking the entire workforce for every question. That's what keeps a model this size affordable to run. Its headline upgrade is quite impressive. A million token window, five times larger than its predecessor, that's large enough to

hold an entire mid-sized code base in the working memory at once. On z.ai's own published benchmarks. GLM 5.2 scored 62.1 on SWEbench Pro, ahead of GPT 5.5's 58.6 and 81.0 on terminal, Bench 2.1, four points behind Cloud Office, 4.8's 85.0. The more interesting evidence comes from two platforms that run blind crowdsourced comparisons. These are ones where users submit prompts,

get responses from anonymized models, and vote on which is better. On designerina, GLM 5.2 actually edges out Fable 5, 1,360 to 1,350. On arena, Fable 5 leads clearly, 1,654 to 1,595. It gap wide enough that the two models statistical margins don't overlap. But looking at the leaderboard, it is impressive of how GLM 5.2 actually beats all the other frontier models at front end coding. It beats all the cloud opus, 4.7 and 4.8 default models and thinking models across the board. And what's more clear here is that the gap between Fable 5 and its other models are quite large, but GLM 5.2 fits right in between where that gap is. However, where GLM 5.2 wins without argument is price.

$3,682 million output tokens at GLM 5.2's rate. That same money buys 200 million tokens of Claude Sonic, 120 million of Claude Opus, and only 60 million of Claude Fable 5. Fable 5 is the most expensive model on that list by a wide margin. within 11 times the cost of GLM 5.2 for the same volume of output. The benchmarks are why developers pay attention to it, but the positioning of the model is why everyone else did. Z.AI released GLM 5.2's full weights under an MIT license. The most permissive open source license available.

Anyone can download, modify, and commercially deploy the model with no fee and no restriction. The company was explicit that this was a response to the export band, not just a coincidence of timing. Its launch message read, intelligence should be open, accessible, and ready to build with, empowering every developer, everywhere. The contrast didn't need to be spelled out. A US government order had just made a frontier model legally unavailable to everyone. The first time this has ever happened and because of it, it may happen more in the future. Z.ai came in just 24 hours later, allowing everyone to have access to a model just as capable

with the addition of it being open source. Once weights are public, there's no central figure to take it down, and it isn't a strategy unique to Z.ai. Deepseek built global attention the same way in 2025, undercutting Western pricing, while open sourcing its weights. Ali Baba's quen and moon shots kimi have followed the same playbook since. Chinese labs have increasingly treated openness itself as the product, a way to compete on access and distribution while closing the gap on raw capability. The export ban on Fable 5 didn't create that strategy.

It handed Chinese AI labs the best advertisement it could have asked for. Now none of that means the positioning matches the reality for most of the people that z.ai says it's empowering. Running GLM 5.2 at full precision on your own hardware takes roughly 1.5 terabytes of GPU memory. In practice, that's 8 Nvidia H200 chips working in parallel. That's out of reach for almost any individual developer. For the large majority of users, open-impractice means routing requests through z.ais hosted API rather than self-hosting the weights.

The accessibility argument is real for a small number of well-resourced teams. For everyone else, it's a trade-off between providers. The benchmark picture carries its own asterisk as well. A 2025 academic paper found that large labs could privately test multiple versions of a model before release and only publish the best performing one. Meta reportedly tested 27 internal variants of Lama 4 before choosing which one to put on the leaderboard. Arena's leadership disputed the paper's conclusions, but its reasonable evidence to treat any single leaderboard

number as a data point, not a verdict. There's an irony on the American side worth stating plainly. Anthropics spent its Fable 5 launch week explaining why the model was powerful enough to need strict safeguards. The government took that admission at its word and applied a far broader restriction than anthropic expected. A company that built its brand on describing its own products as almost two dangers to release had written the justification a regulator would later use to shut one down. Now this could be a more political reason where Donald Trump just wants to mess with anthropic

because of past events. That is up to discussion. Looking at this holistically, an export control aimed at one jailbreak ended up demonstrating something else entirely. Restricting access to a model doesn't restrict the competition driving in. Anthropics still has the better model for the most part, but it doesn't have a meaningful lead in price, in licensing flexibility, or in how convincingly it can tell a story about access. Z.A.I. didn't need to win the benchmark.

It needed the ban to exist and a day to act before the news moved past it. That's the part worth watching going forward. Every time a Western lab treats its own model as powerful enough to require government-level restriction, it hands Chinese labs a ready-made argument. That theirs is available and worth using. This argument won't go away after Fable 5 is released back to the public. just waits for the next company to give it a reason to make the same case again.