alt Hacker NewsGLM 5.2 and the coming AI margin collapse
Comments
They have a vision MCP to make up for the model itself not having the capability natively: https://docs.z.ai/devpack/mcp/vision-mcp-server
I also found their web search to be mostly okay.
Furthermore, in case this is of interest to anyone, if you use their ZCode harness then you get bigger Coding Plan quotas: https://zcode.z.ai/en
Used it for a bit, it sits somewhere between OpenCode Desktop (still new but nice) and Claude Desktop (recent versions are good).
As for GLM 5.2 as a model - with max thinking it’s generally satisfactory, somewhere between Sonnet 5 and Opus 4.8, better than DeepSeek V4 Pro for sure.
Pricing wise, the subscription doesn’t seem as good as expected. I spent like 60% of the weekly limits of the Pro (50 USD) plan in one day, only because each 5 hour limit only gave me 20% to spend, otherwise it’d be 80-100%. Not even doing anything crazy, just parallel long form work on 2 projects with about 96% cache rate and at most 3 parallel code review sub-agents.
Their Max (100 USD) subscription would last me the whole week, but so does Anthropic for the same money and so would OpenAI. Off-peak is more palatable but I can’t just twiddle my thumbs at 9 AM to 1 PM local time.
Proper savings would show up with the Max plan and yearly billing, but that’s more of a tough sell.
I'll agree but from the other direction. AI continues to absorb my job as a senior systems software engineer (c/c++) and after a couple months I've only spent a few hundred dollars using gpt-5.5/5.6 and codex. I have no idea what people are doing to burn so many tokens but for me this is laughably cheap and every day I discover new capabilities. I don't care if costs go up or down, it's so cheap for what I get that I don't care.
It’s important that none of these entities can collude to price fix. Having China be the competitor ensures that.
Basic microeconomics is still the easiest way to understand token economies. How is it not a competitive market (where profits go to zero?).
Anything A or O does to keep more margin, any competitor can copy or choose to undercut, and undercutting has the benefit of collecting training data. So what is going to stop gross profit of tokens going to zero except for collusion/price fixing?
> Of course, this was a hugely poor read of where the costs actually lie in AI. Training - while no doubt capex intensive - is a fixed, up-front cost. You spend hundreds of millions to train a model, then you are "done".
I don't understand this point that people make. If you're consistently needing[0] to train new models and the cost of training relative to the % improvement seems to go higher, isn't this just a constant cost that you continue to bear? The footnote seems to allude to this, but then sort of waves it away anyways. Also are there continuing incremental training costs to keep models relevant? Or do they only have knowledge of events up to the day they were trained?
[0] needing, because you have competitors and people expect more and more.
Metaphor i like is that it will be as cheap as electricty?
Do you know who is supplying your electricity or which factory it runs on? probably no, bc its a commodity and mostly settled and there is so many energy resources. some are alternative some are coal mines. And they all fight in the supply demand trade for energy which is happening real time ( think open router here)
And eventually the consumer wins bc of the abundance.
I think greatest example of abundance of cheap infinite intelligence will be not glm5.2 but DeepSeek V4 Pro max with $0.435 per 1M input tokens and $0.87 per 1M output tokens
>the least understood upcoming shift in AI economics.
Then proceeds to talk about something in the AI news every day. Hey, did you guys hear? Open source models are cheaper and their quality is increasing!
So, first, by no measure is GLM5.2 as good as Opus.
Second, yes, open source models will put pressure on margins...eventually. Everyone knows that. But do you think today's AI business model is the same as tomorrow's?
Seems like a pretty pointless post that still centers around output tokens.
In agentic coding, cached input tokens is 90% of the API "cost". It doesn't require GPU compute, and DeepSeek has shown that it can be done 50~100x cheaper with MLA/CSA/HCA, and a whole bunch of disks. This should collapse the margin.
I also think we’ll approach a point where increasing intelligence is not really going to suddenly improve most work tasks. I bet that’s already happened actually.
We’re oohing and aahing about models, when the ones a few versions ago did a good enough job for most of the dumb coding, etc we do
Braintrust which is a really solid eval tool/platform just compared it to Opus 4.8 to see if it could preserve exact long context retrieval under prod serving constraints and it did really well. I think 6-12 months before OSS has Fable-esque models
> I'd be very surprised if it wasn't more than 50% cheaper for nearly all workflows, for a very similar level of quality.
If your using pure API ... providers like neuralwatt cut that cost down even more by using energy as the actual cost. So GLM 5.2 is more expensive then GLM 5.1 on their service (those thinking tokens), compared to API costs, its dirt cheap. And way more tokens then the zai subscription delivers.
We are seeing a move towards more realistic pricing on actual consumption based usage. Be it DeepSeek, Xiaomi (MiMo), or zai's GLM via neuralwatt.
The main issue facing subscriptions a-la-carte usage, is that a lot of the heavy hitters really drain the resources. And that as a business model can not survive without ...
a) increasing the prices. b) everything goes to actual token/energy usage based billing but with more realistic pricing, and not the bloated API prices that are focused on companies.
We shall see what the future holds but things will change.
I think the profits depend on how well they manage their fleet purchases (or possible sub-leasing?) to get high utilization without overloading or idle racks.
Because accelerators like H200, B300 etc. are highly parallel and designed to run like 200 or maybe 300 sequences at once (depends on the model, just guessing). I assume they finance the hardware and that cost per device or rack is the same whether each unit is handling 10 requests or 150 requests (aside from electricity).
And probably international customers factor into it to get good utilization over more of the night time. And it likely is something that they look at quarterly more seriously than monthly. The biggest risk to profits might be a downturn in business that causes some portion of the financed AI accelerators to go idle or get low utilization for some weeks (that they can't sublease).
IMHO, cheaper inference means higher costs overall :) because everyone will use more thus driving up the investment required to stay current or to compete.
Switching models is also kind of easy but not plug-and-play. Most harnesses out there do very poor job with the open weight models. Unlike Opus, GLM 5.2 ends up in loops and hallucinates a lot more. If your harness is built on the expectation that the LLM will perform well, then switching to GLM 5.2 will be an uphill struggle. We had to refactor our harness and introduce more defences because of GLM.
The cost savings are substantial. Obviously it really depends on your workloads but it is noticeable cheaper for agentic work. Coding - I don't know. We do have some coding agents on GLM 5.2 and what I noticed with some landing page experiments that the results between GLM and Opus are identical - they might be using the same training data? Obviously Opus is still substantially better model. I don't think there is an argument to be made here but GLM 5.2 is cost effective and really good too.
Overall, we switched all of our internal agents to GLM 5.2 and because it is Open Weight we are in talks to get the model from certain geo locations giving us more freedom as well as extra protection.
Overall I think this industry will be in much better place because of GLM 5.2 and whatever open-weight models come next.
> It turns out that nearly every agentic session does a lot of web searching for looking up items
This is why Google will win the race over most of its competitors. They own search.
How fast is glm 5.2 in western hosts? It's doing everything I want it to, but going through PRC host it takes like 5-10 times longer. Not sure if that is nature of modest or PRC computer infra/routing.
“Z.ai provides a replacement MCP for web search, but it's pretty awful and slow”
I’ve had good results with Tavily so far, might be worth checking as an alternative for agent search.
I don’t understand the argument here. The article doesn’t describe a collapse or the breadcrumbs for it. The only argument I can put together is companies hosting the open source models in house or use some service like Amazon that could potentially host them and so replace the frontier models. Data center and specifically infra to host llms is still the main sticking point given the security concerns about data going to china. The article doesn’t make these arguments coherently
How long will that $4.40 rate persist? Until we know more about the real unit economics it will be damn near impossible to rely on steady inference costs or make them predictable at the enterprise level. Gonna be a wild ride for awhile.
I don't think the writer has used top tier models very much. I have subscriptions to basically every provider, the difference between glm5.2 and opus is not even close, the gap is huge. raw benchmarks glm is impressive , but in practice these models are lacking so much. I had fable create a detailed implementation guide that explained how to implement everything in immense detail, it included all the libraries to use and versions. I then had deepseek v4 pro execute and it used old versions , different libraries and cut corners. Fable said about 80% was implemented wrong.
I had GLM 5.2 do the same, and it performed exceptionally better, but when it got stuck on something it would be trial and error mode going forward and have zero foresight for future issues that might occur due to fixes it was trying. the model severally lacks prompt understanding, and testing .
As long as the SOTA models are 'ahead' then there will be a big premium.
in cursor benchmark glm5.2 is on par with gpt 5.5 medium and sonnet for the same task from results and cost perspective.
The speed of generation for both gpt 5.5 medium and sonnet 5 will be dramatically faster. source : https://cursor.com/evals
I don't get the hype. It's near SOTA model that is not deepseek of this world. It an expensive to run model, and under certain tasks it is comparably cheap as closed source ones.
the economics of this are a little counterintuitive.
is there a market saturation point for intelligence? how about for software? it seems like the more you have the more you want because you're trying to do more things.
as the models get smarter I get busier because I'm doing more things...
‘640kb of ram should be enough for anyone’
Disclosure - Fireworks kindly gave me some free credit to experiment with GLM to help write this article.
GLM is the model that will sink the frontier labs.
Recall last year deepseek? And 18 month's later? What changed?
Is this AI written (or edited)? The word "genuinely" appears 4 times on the page.
Man, I hate how often people/LLMs use that word now. Maybe other people gloss over it but it's super distracting to me.
This article only promises to get into "the coming AI margin collapse" in a yet to be published "part two". This part only makes the point that GLM 5.2 is pretty good (no shit).
I hope cheaper inference eventually means faster speeds at the lower tiers. I don't want to settle for 100 t/s, but I don't want to pay $10 per prompt either
I would not be unsurprised if the US govt steps in to prevent this. They'll do anything to stop China getting ahead in the AI race.
There's the sanctions already implemented, next step might be giving these companies government funding, just like they do with military companies.
how do massively negative margins "collapse"
Inference has been decreasing in cost by about 10x per year since 2023.
Yes, margin on model inference is high with some providers. If you just wanted inference (at cost), you'd buy a GPU, or rent one from AWS or Microsoft. But you're not paying OpenAI/Anthropic for inference. You're paying them for a platform. Every feature OpenAI/Anthropic bake into their applications, models, online services, etc - anything that isn't pure LLM text generation - is a custom integrated add-on service that LLM weights do not include. Even if open weights became cheaper and better than OpenAI/Anthropic, most people would still pay for OpenAI/Anthropic, because they give you things the weights alone don't give you.
Comparing Z.ai GLM 5.2 to Claude Code w/Opus 4.8 is like comparing Linux Kernel 7.0 to Microsoft Windows 11. If you don't know much about computers, you'd say these are the same things. If you know a lot about computers, you know the latter has a thousand extra things that make a huge difference in what it does out of the box. Which one you use speaks to what kind of customer you are.
Sure, GLM 5.2 doesn't have vision; but an AI power user can plumb together any VLM with the text generation of GLM 5.2 in most AI harnesses, just like a Linux power user can combine the Linux kernel with KDE Desktop. Most people don't use Linux and KDE, because it's unpopular, difficult to use, hard to get support for. Instead they pay for Windows or Mac, because there's lots of support, with a giant company pouring money and effort into filling all the usability gaps, making it seamless.
Most people don't pay for the cheapest possible thing. They pay for the thing they can afford that improves their life while making it easier. An open weight alone is almost completely unusable by itself (like the Linux kernel), compared to an AI platform (a completely usable system). If you're constantly wondering about when open weights will reach parity with OpenAI/Anthropic, you're a Linux person. If you just pay $20/$50/$100 for OpenAI/Anthropic without thinking about it, you're a Windows/Mac person. There is nothing wrong with either of these groups, but they are fundamentally different, and always will be. An LLM weight is simply a different category of thing than an entire AI platform/provider.
> Where it gets really scary for the frontier labs is how easy it is to migrate to open weights models. Both Z.ai and Fireworks offer both an OpenAI compatible and Anthropic compatible endpoint. This makes it absolutely trivial to use with Claude Code and Codex.
Yes the ease of switching is greatly appreciated.
Now the reason I tolerate Claude Code in my tmux sessions is because apparently Anthropic ain't playing nice with the subscription plans and other harnesses.
But I'm evaluating pi.dev atm and it looks amazing. To me being able to rid of that piece of vibe-coded underperforming, characters-modifying, turd that Claude Code is a big motivation to switch to GLM (I'll probably keep my OpenAI subscription as OpenAI repeatedly said they were cool with other harnesses).
It's also quite obvious that Claude Code is receiving new vibe-coded slop features after vibe-coded slop features in an attempt to lock you in.
To anyone thinking about switching to GLM: I'd say at least evaluate pi.dev and see if that wouldn't be an opportunity to kiss Claude Code and its "gameloop that converts characters from a headless browser to other characters to show in a terminal at 60 fps" goodbye once and for all.
I think OpenAI, Anthropic and SpaceX are going to envy the dinosaurs because there's not asteroid coming for them, there's three:
1. There will be no moat around frontier AI models in the future. China is going to make sure that happens. It's a national security interest for them. DeepSeek was the first shot across the bow for that but it won't end with them. There are other labs and there are non-Chinese actors too. The stratospheric valuations depend on there being that moat; and
2. Nobody seems to be considering what the next generation of AI hardware is going to do with current hyperscalar investments. We're about to go through this with the B100/200 move to R100/200 but a lot of the investments are probably slated for that next-gen. But what about 3 years from now when the hypothetical X100/200 comes out and doubles FLOPS and halves performance-per-watt. What will that do to existing investments? Some people are delusional and think that they'll get 10 years out of GPUs when 10 year old GPUs (eg V100) are sold for scrap and 5 year old GPUs (A100) cannot run DeepSeek v4 Pro. And people think the A100 is going to get another 5 years of use? No; and
3. Local LLMs are coming for remote usage. You can buy a 5090 PC for less than $5000 currently but you're limited to 32GB of VRAM, which will comfortably run 31B models but nothing really larger. Go to $12-13k to upgrade to an RTX 8000 Pro and you have 96GB of VRAM, which will run larger models (but certainly not, say, DS v4 Pro or even Flash). You have shared video memory products rapidly coming from NVidia's aggressive market segmentation. Things like Strix Halo and DGX Spark have severe limits on memory bandwidth (<300GB/s compared to 1.8TB/s for a 5090/6000 Pro and 3TB/s+ for server grade HBM3e/4 based GPUs). Macs could be real interesting in this space butr they lack the raw FLOPS with the M5 generation.
But what will this local hardware look like in 2-3 years? I think people will be shocked at how much better it will be with the Apple M7 Pro/Max generation (2028 expected) and the RTX 6000 cards at that time although I fully expect NVidia consumer GPUs to still top out at 32GB of VRAM to maintain that segmentation. And I look forward to what the next generation of the AMD Ryzen AI Halo platform will look like if they really try.
All of this adds up to these three companies needing to cash out before the music stops (IMHO).
[dead]
i would use glm 5.2 if the servers weren't in china
i mean i guess my employers wouldn't know the difference
but i'd like to play it safe and keep everything in america
The fact that these Chinese models are getting close to “Opus-grade” despite costing 6x-8x less is huge.
As the token bills start to come in, those economics will be harder to ignore (regardless of the origin of the LLM); especially as there will be many CIOs sweating over their quick and costly AI initiatives showing little ROI.
My hope is that the EU also steps up their own competition in the frontier model space so that it’s not just China v USA.
I'm not convinced raw costs matter:
1. Compute costs collapsed since the advent of Cloud and yet hyperscalers still have fat margins.
2. Many open source office suites exist yet none compete with the ubiquity of gsuite or office. GitHub, Slack are similar examples.
3. Both Windows and macOS dominate the home desktop space despite free alternatives existing for a long time.
4. Many formerly open source infrastructure components like Redis and Elastic Search have Apache equivalents, but they still command healthy margins.
I understand the arguments for a margin collapse, but I don't see any historical analogues. It seems that enterprises will pay top dollar for service guarantees, integration, and someone they can sue.
It's nobody gets fired for buying IBM all over again.