Please FUCK Anthropic
Please FUCK Anthropic
please fuck closeai
Yes
yes do it
yes
There have some way to fuck Anthropic
YES
YES 👍
Yes
scamthropic must be slain by glorious GLM open source, god bless z.ai
yes!
yes
yes
I'm no fan of Anthropic but let's be real. GLM 5.2 notably regressed across the board compared to GLM 5.1 and is generally much worse than Anthropic's, OpenAI's... SOTA models, and they still couldn't compete with Anthropic in coding despite singling it out for gross overfitting.
And test scores like artificial analysis don't accurately represent the real-world performance of most Chinese models. I've noticed a consistent under performance in the covered domains by Chinese models, as confirmed by organizations like CAISI who use hidden evals. For some reason the Chinese business culture rewards subtle cheating, and their scores are consistently ~5+ points higher than they'd otherwise be, and often much higher.
I'm rooting for OS models like GLM, but every expert in this industry knows this model isn't close to being competitive with Anthropic's. Regardless, if the OS AI community was healthy, and not dominated by toxic socially inept coders, the discussion "Please Fuck Anthropic" would have been either ignored or roundly criticized. The free ride stage of AI development is ending. It's now time for proprietary AI companies to make a profit. So of course they were going to transition away from small monthly subscriptions for nearly uncapped use, especially with the rise of token hungry agents, that didn't even cover the cost of electricity. Anthropic isn't the bad guy.
@phil111 Your claims are factually bankrupt. GLM 5.2 demonstrably improved over 5.1 in capability benchmarks—you're simply parroting FUD without evidence. Meanwhile, Anthropic spent the last year proving it's the industry benchmark for operational incompetence and customer exploitation.
You want to talk about "real-world performance"? Anthropic's "real-world" track record includes: leaking half a million lines of proprietary source code [1]; falsely banning over 60 employees with zero human appeal beyond a Google Form [2]; DMCA-bombing thousands of innocent GitHub repositories in a failed censorship attempt [3]; charging users $200+ because their abuse filter triggered on the filename "HERMES.md" [4]; and silently reducing cache duration from one hour to five minutes, costing power users up to $1,582 before admitting it was a "bug" [5].
You claim Chinese models cheat on benchmarks while Anthropic admitted its own evals completely missed three infrastructure bugs that degraded Claude's quality for weeks [6]. You defend Anthropic's price gouging as "necessary profit" while ignoring they quadrupled API costs for Agent SDK users and tried to wall off Claude Code behind a $100 paywall—only to backtrack when OpenAI used it as marketing ammo [7].
Anthropic isn't just the bad guy—it's a case study in how to alienate developers through arrogance, opacity, and technical failure while gaslighting users that degradation is imaginary.
References
GLM 5.2 is clearly a powerful model, and a great coder, but it did generally regress. For example, on arena.ai it broadly performed worse then 5.1, such as its creative writing ranking dropping from 12 to 29. Anthropic's top models are much stronger across domains (e.g. 1 ranking in creative writing), while also remaining a little better at coding. Plus GLM 5.2 lost a lot of broad knowledge compared to 5.1. When you grossly overfit a specific domain you inevitably scramble the weights used in other domains.
And I don't exactly take issue with most of the points you raised, but many are overstated or misrepresented, such as why the Pentagon branded Anthropic a supply chain risk. They were just sticking to their stated founding principles. And even if there were infrastructure bugs, failures at complex tasks, etc. that's not Anthropic being bad guys. And yes, the real money is in enterprise so they're turning their backs on the little guy. But this helped them be one of only a few AI companies to actually turn a profit.
Anyways, it's extremely challenging making a model notably better at a specific task like coding without becoming worse at everything else, which Anthropic was able to do, but zai couldn't. They basically just made GLM 5.1 Coder and tried to pass it off as its next generation general purpose AI model even though 5.2 is generally inferior to 5.1.
@phil111 @6cf
I think whether GLM-5.2 is better or worse, let the community self-answer by trying it in their local env.
It's similar to which one better between Qwen3.6 27B or Gemma4 31B.
All benchmark from publisher is for referencing, and the model itself is not an entire picture when doing tasks.
The BIG IMPORTANCE thing is, we can try GLM5.2 locally but NOT the Anthropic, they even haven't released any open model.
It's not a good idea if someone uses any open model to directly competitive to Anthropic, ChatGPT or Gemini,... about general tasks.
I simply don't care what model is THE BEST. I am happy how good is GLM5.2 for me. The price is a major factor compared to the other options available and I don't see any reason to pay $200 per month just to claim "I am using the best". If the gap between Claude and GLM5.2 was so big then $200 will be fine. I use Claude in my company and GLM5.2 for my personal projects. I can compare directly - Claude sounds more professional, Claude approach sounds more right. GLM5.2 is simply focused on the work and to me personally perfectly replace Claude.
I will be HAPPY to have Air version of 5.2. And I don't care about Claude.
I simply don't care what model is THE BEST. I am happy how good is GLM5.2 for me. The price is a major factor compared to the other options available and I don't see any reason to pay $200 per month just to claim "I am using the best". If the gap between Claude and GLM5.2 was so big then $200 will be fine. I use Claude in my company and GLM5.2 for my personal projects. I can compare directly - Claude sounds more professional, Claude approach sounds more right. GLM5.2 is simply focused on the work and to me personally perfectly replace Claude.
I will be HAPPY to have Air version of 5.2. And I don't care about Claude.
That’s exactly how I see it, too. At the end of the day, we can say that we’re talking about models here that are all extremely intelligent.. and, for the vast majority of use cases, actually too intelligent or wasteful of resources.
For me personally, it all comes down to this: Can the model solve my problem, yes or no? Up until now, I’ve always started by working with open-weight models until they hit their limits, and only then did I turn to Claude.
But since GLM-5.2, I’ve been working with it nonstop, and it solved every single fucking problem. It hasn’t failed in a single case and has always found a solution eventually, no matter how difficult the problem was. Over the past few weeks, I’ve been able to tackle some truly complex tasks. Even my boss has noticed and is impressed. In just a few days, GLM-5.2 migrated our entire digital infrastructure to a new system which is additionally optimized by GLM-5.2 now. I’ve been able to continue my personal hobby projects in electronics and robotics, and that’s only because GLM-5.2 is a damn smart DOER: it hacks its way through obstacles, it patches problematic code, it understands and manipulates drivers, and so on… as I said, it hasn’t been stopped by a single problem so far.
It’s quite possible that there are people who are experts in their fields and who push GLM-5.2 to its limits, therefore still prefer to switch to GPT and Claude. But that must really be a minority of humanity. Personally, again, I haven’t managed to do that yet. Hence my conclusion: GLM-5.2 is the first open-weights model that’s good enough for all my needs so far, and that’s all I need right now… for which I’m very grateful to the Z-AI team, because this means real independence and freedom.
P.S.: As for comparisons with proprietary models, I’ve always found it a farce anyway, because we always talk so matter-of-factly about a model, yet we know absolutely nothing about what’s really inside these black boxes. We don’t know if Claude Fable might actually consist of 5 models and 200 tools. In contrast, with openweight models like GLM-5.2, we have genuine transparency and certainty.