GLM 5.2 just did what no Chinese open-weight model has done before — it survived first contact with real usage. As power users rave and a Nobel laureate flees DeepMind, NLW separates fact from innuendo on the Fable saga and asks whether the two-horse race is truly over.
GLM 5.2's DeepSeek moment shatters the two-horse race — for real this time.
Chinese open-weight models always benchmark well and then fade fast. GLM 5.2 is different: respected practitioners are raving days later, it beat Fable 5 on website design, and it's forcing companies rethinking their stack amid the Fable ban and compute crunch to take open models seriously. The lesson isn't to race out and buy GPUs — it's that the flowering of diverse model architectures means every organization should have a sandbox to experiment with the alternatives.
Last week's first impressions of GLM 5.2 were good. After a weekend of hands-on usage, belief in the model and its implications has done nothing but grow — landing as the Fable ban and compute shortage already had companies rethinking whether to fire up the most state-of-the-art model for every use case.
Practitioners are calling GLM 5.2 a turning point as significant as DeepSeek R1, with one noting an open-source model broke into the top three coding models faster than anyone expected. The original R1 moment came when DeepSeek put a free reasoning model in front of casual users — and peeled $589B off Nvidia in a single day before the panic receded.
DeepSeek's weird legacy: every new Chinese open-weight model scores high on benchmarks, everyone says the gap has closed, then a couple weeks later no one's using it. They tend to fade almost instantly — which is exactly what makes GLM 5.2's staying power different.
As the overall state-of-the-art rises, being three or six months behind has a lot more viable use cases than it did a year ago. Chinese models have steadily been integrated into the stack, especially for startups and smaller companies with fewer constraints on which models they can use.
— Itamar Golan. It's not just hypebeasts: Vercel's Rauch was 'almost shocked' at GLM 5.2's coding, and Itamar Golan said that for the first time an open model felt meaningfully close to frontier-lab quality across real tasks. The vibes have been clearly different from past Chinese releases.
Per Design Arena, GLM 5.2 ranks first on websites specifically, though it trails Fable 5 on game development, data visualization and 3D design, and sits fourth on UI components. Three behaviors drive it: better starting templates that avoid AI anti-patterns like purple gradients, natural use of libraries like Chart.js and Three.js, and more detailed outputs.
— Theo, AI entrepreneur. Contrary to the assumption that Chinese open models are dramatically cheaper, YouTuber Theo notes both Opus 4-8 and GPT 5.5 (medium) are cheaper and smarter than GLM 5.2. The tokens cost less, but the sheer volume — GLM produces ~25% more code and takes about double the generation time — means you spend more and wait longer.
Some are acting like the only way to run GLM 5.2 is locally — Itamar Golan estimates ~$400K to buy or ~$20K/month to rent the eight H200s needed. NLW's recommendation: rather than hacking complex physical infrastructure, just try it via a routing tool like OpenRouter or an open-source harness.
— Elon Musk. Debating when China gets a full Mythos-class model, Elon Musk argued Q1 — pushing back on the ZAI founder by drawing a line between benchmark performance and true usefulness, which 'does not show up in benchmarks, but definitely shows up in revenue.'
— Aaron Levie, Box. Box's Aaron Levie argues credible frontier-level open-weight models should be a huge update: they let you guarantee sovereign AI, post-train for specific workflows, cost-optimize across workloads, and afford to do much more — a 'huge win for the applied AI layer.'
NLW's bottom line: even if Fable returns this week, the idea that AI is a two-horse race between OpenAI and Anthropic (with an asterisk for Google) is over. Intensifying, costlier workloads plus government review of the most powerful models plus viable runner-up models point to a flowering of diverse architectures companies can optimize for speed, cost or performance.
NLW doesn't think most companies need to race off their core subscriptions. But having some part of the organization with license to experiment with alternative model architectures is probably time and money well spent right now.
A resurfaced Economist line quoting Senator Mark Warner saying Mythos broke into almost all classified systems 'not in weeks, but in hours' lit up X. But the reporter, Shashank Joshi, said it would be a mistake to read it literally — it described Mythos's potency under very particular controlled conditions, not a real breach.
— Peter Wildeford, AI policy commentator. AI policy commentator Peter Wildeford offered more plausible readings: a simulated exercise against replica systems, Mythos given code and docs up front, poorly secured internal IT mislabeled as classified, or heavy human tooling. His point — none of it makes the underlying cyber capability less alarming.
Additional reporting from CyberSec Guru confirmed the incident occurred during an NSA red team exercise — not an outside attack. They also noted NSA Director Rudd is a contested new appointee whose background is special operations, not signals intelligence or cyber.
— President Trump, on the Axios Show. Asked on the Axios Show whether he regards Anthropic and Dario Amodei as a national security threat, Trump struck a conciliatory tone — ruling out the Defense Production Act, saying 'we're beating China by a lot,' and praising Anthropic for responding 'very responsibly so far.'
AlphaFold creator and 2024 Nobel laureate John Jumper left Google DeepMind for Anthropic — the second high-profile departure that week after Noam Shazeer left for OpenAI. Some suspect being reassigned to AI coding rather than AI-for-science contributed to his exit.
— DeepMind source, via Leo at SynthWave. A DeepMind source told reporter Leo at SynthWave that morale is cratering as staff perceive the lab slipping to third or fourth place, demoralized by ZAI's GLM 5.2 overtaking Gemini 3.1 Pro. Gemini 3.5 Pro — called 'not the step change we need' by another source — is reportedly slated for June 30th.
NLW cautions that behind-the-scenes reporting needs a grain of salt and that we tend to make too much of individual departures — humans have complex motivations. But two high-profile DeepMind exits in a week does start a pattern, and Google's slide in the coding and enterprise race is genuinely notable.
— Andrew Curran. Andrew Curran reports a more capable version of Mythos has emerged from training, and notes that embargoing a public model does nothing to slow development — it may even speed it up by freeing resources. Labs can't afford to pause, and GLM 5.2 is proof of why.
The slug 'Claude Sonnet 5' appeared on an Anthropic partner provider, and some report GPT 5.6 already showing up in Codex. Speculation runs to a simultaneous Fable 5 return plus Sonnet 5 plus GPT 5.6 all landing in the same window.
— Thibault, OpenAI Codex lead. OpenAI Codex lead Thibault began vague-posting about dramatically better front-end abilities in upcoming models. Scientist Derya Nutmaz, who often gets early access, added that those who think Fable 5 will stay the best 'will soon be proven wrong' — stressing 'soon,' not 'eventually.'
