Why aren't more people talking about this? It's literally Opus 4.7 quality stupid prices. I know providers who are offering this at unlimited tokens for $50 a month. Some are even offering API rates at 3x lower than the official ZAI api rates which are already like 10x cheaper than Opus. (Crof and Umans btw)
This is a huge blow to Anthropic/OpenAI/Google and a massive win for the rest of the world. The official API prices and speeds mean nothing for open source models.
Be careful about unofficial providers, a lot of them misconfigure models or stealth quantize them. For a while the difference between Kimi on the official API and most third party providers was 20-40%.
We use three classes of signals:
* Tool-calling success and reliability from real traffic
* Provider performance metrics such as throughput and latency
* Benchmark and evaluation data as it becomes available
> Some are even offering API rates at 3x lower than the official ZAI api rates
Looking at openrouter [1], some of the cheaper offerings are for quantized models. Not sure how much intelligence is lost in quantization. And they are not 3 times cheaper. Where did you find 3x lower prices for APIs? I am considering skipping open router and using them directly for that price.
Neuralwatt ... When you reverse calculate the actual energy usage / price on a token basis, the gap is large.
I do not have GLM 5.2 numbers because the whole default max setting is overkill. But GLM 5.1 numbers had it at 12x cheaper then API rates. And about 2.5x more tokens vs zai their own subscription service.
Yes, its FP8 but lets be honest, do we know for sure that even zai runs at FP16? I learned a long time ago with Claude and Codex how much cheating happens on model levels, even from the big boys.
Please correct me if you have contradicting data but: Neuralwatt's price per token vs price for energy comparison doesn't seem to take into account the cost savings from cache hits that other providers offer on pure token rates. The comparison seems to assume every input token is a cache miss.
On top of that, the cloud offering doesn't seem that well-run, they randomly blocked a colleague's API key for a couple days without any heads up, had a weird rate limiting bug and they have been deprecating models without redirects with very short notice, all while taking weeks to onboard new models. I assume some of these problems would be addressed if we had an SLA/enterprise contract.
It's a promising idea though. They offer a $5 trial credit (with an aggressive rate limit) though so no harm in trying it out.
> doesn't seem to take into account the cost savings from cache hits
Absolute false information.
From my usage panel for this month:
* Total Tokens 1.1B
* Cached Tokens 1.0B 97% of prompt tokens
* Cost energy pricing $26.58
The energy pricing is higher then what i actually pay because its a mix of token billing and partial subscription (60% extra "power").
From the $50 subscription, i have about 3/4 left (4.21 of 16.0 kWh used this billing cycle). Used $5.5 in token billing.
That was running 82.0% GLM 5.1, and 18% GLM 5.2. Yes, i have been busy ;)
My actual usage if we look in dollar value was ~ $18.
For your information, that is cheaper the MiMo v2.5 Pro from Xiaomi as there i was doing around 450.000t per cent. And they have the same 75% cheaper prices like DeepSeek. MiMo has a issue with cache retention between session prompts what hurts them vs DeepSeek. Yes, DeepSeek v4 Pro is 2.5x cheaper but nowhere near GLM 5.1, and especially not GLM 5.2.
In case your wondering, zai subscription light is about 80m token / week limit. So on a token/cent price, neutralwatt is about 3x cheaper (and not 5h, week limits to maximize/frustrate).
> all while taking weeks to onboard new models.
Took them 1 day to include GLM 5.2 ... Yes, the remove old models fast because they do not have the server capacity to keep old models around.
> I assume some of these problems would be addressed if we had an SLA/enterprise contract.
Its a small team, not a big huge company. From my experience so far, seen a 2 timeouts, and sometimes slow speeds as servers get overloaded. For what i am paying for GLM ~5.1~ 5.2 ...
Your reply doesn't seem to be in good faith. Please provide your formula for calculating effective per token cost.
I am not sure why the small team argument is relevant. This is a crowded market, there are dozens if hundreds of third party inference providers in the world right now. I'm glad that's a good excuse that works on you but I'm not sure why the average user should care.
The formula is very easy. Go to the website of neuralwatt, and read ... 5$ = 1Kwh in power for non-subscription usage. For subscription usage you get ~50% more.
Then you actually use the service and see how much tokens you use on average. You calculate the token use vs what you pay. And this gives you a stable number to compare different services and model with, if you want the token cost. This is basic school level reasoning and calculation.
> I am not sure why the small team argument is relevant.
This is relevant to the previous poster his question regarding support and SLA/enterprise support.
> Your reply doesn't seem to be in good faith.... I'm glad that's a good excuse that works on you ...
Question: Do you have a issue with communicating with other people in real life?
IME, unquantised -> FP8 is pretty much lossless. What matters more is having an unquantized KV cache - using an FP8 KV cache can result in a significant drop in quality.
I don't know, sounds quite similar to his rate distortion theorem (analyzing minimum number of bits/symbol you need to stay under some fixed amount of distortion). I.e. lossy compression with a maximum amount of loss. I.e. "pretty much lossless" compression.
"Pretty much" doing a lot of work. But it's kinda analogous to 99% JPEG compression: yes you can detect loss, but you get meaningful compression ratios out of it and the subjective appearance is nigh-on perfect.
Shannon would be pointing out that if you can throw away half the model without apparent degradation, we're nowhere near packing in all the information we could in training. There must be a better arrangement than we've currently got.
I've seen a few articles from providers talking about KV cache quantisation, but it's not something they explicitly point out like they do with weights.
So you could end up paying more for unquantised weights, only to get silently hit with a quantised KV cache...
To answer the question in your first sentence - because it's VERY computationally (ha) expensive as a human being to keep up with all the options. It's also very hard to figure out how to run a model like this. There's no installer. If you really really care, which 99% of people do not, you have to google a guide, and then find out it's out of date...
I've tried a number of these, and the learning curve is very steep compared to "install Claude Code and pay $100/mo". There is no way saving me $50/month matters compared to figuring that out.
You're seriously suggesting that setting up opencode or tweaking your claude code config or etc is too much trouble to be worth saving $50 /mo? That's absurd. Doubly so when the audience in question is already using LLMs so ... just ask your existing LLM for help if it seems daunting.
I'm not just suggesting that, I'm trying to be crystal clear: it's a gap that probably cuts TAM by 95% or more. Most LLM users are not software engineers. Even those that are don't care enough to muck with their settings to try out a model. Keep in mind I'm not answering the question "Is this hard to install?" - I'm answering the question "Why aren't people talking about this?"
I would broadly agree with this (based on years of dealing directly with user-facing UX and setup steps). Small hurdles, even easy ones, create larger barriers to adoption then you’d think.
For me it's about tolerance. When I was 13, I could and would customize everything, so much that the computer repair shop told my father that their son "likely is a hacker or something".
At 40, I could easily configure claude code to use another model, even if there weren't any official guides with a bit of MITM fun, but I don't want to invest my attention / heavily use something that will most likely break in the near future.
Here are a few frictions I see that reduce reach, in order:
1) You haven't even heard of it.
2) You have to know to look for both GLM and Z.ai. These are usually in the same article when reporting about GLM is written, at least.
3) You have to understand there could be a benefit in trying it; you have to want to try it for some reason. Their own blog post puts it below Opus 4.8 in each of the three benchmarks they used.
4) You have to figure out the pricing, which isn't obviously in the blog post...
5) When I first went to Z.ai, I got an error popup (not logged in): "You do not have permission to access this resource. Please contact your administrator for assistance." I am using a personal computer...
6) When I typed something in the resultant field and pressed enter, I got "Clear Current Chat? To start a new chat, your current conversation will be discarded. Sign in to save chats"
I think today's article helped with 1 and 2, which helps their top of funnel. But they're fighting a big uphill battle.
It's also very hard to figure out how to run a model like this. There's no installer.
Yes, there is. It's called Claude Code. Point it at the HuggingFace URL and say "Download these weights and build whatever is needed to run them, then test the model."
I really miss the time when people thought that the idea of someone telling an un-sandboxed AI "do whatever is needed to X" was unrealistically stupid.
(In all seriousness, I agree this is a problem. That capability is too powerful not to take advantage of, though. Nobody needs to struggle with this sort of thing anymore, but yes, obviously, it should happen in a VM or at least a container.)
In my org everyone is extremely Claude-pilled to the point you’d think it’s the only LLM that exists, purely because it caters to non-engineers within enterprises.
Wasn't this released like 2 days ago? Everyone is still evaluating and playing around with it, things like the submission is just starting to come out. Give it some days at least before jumping to conclusions, ideally weeks.
I've tried Chinese open models few times before. They were fine, but they didn't come close to the benchmarks they were claiming.
Now, maybe GLM 5.2 is close to Opus 4.7, but I don't wanna keep checking them and keep finding that they're still benchmaxing and aren't at GPT (my choice) or Opus level. The boy who cried wolf, I guess.
Yes, my experience has been the same as yours. I find that the performance of open models is quite acceptable, even good, at one-off questions or small tasks. But they are quite unreliable at long horizon goals.
OpenRouter has a list of providers, looks like NovitaAI would meet those criteria. Though not for $50/mth for 80/M tokens, which I assume is the Z.ai subscription pricing.
The Chinese open weight models have been ahead of Sonnet (at least for coding) for a couple months now. I tend to take benchmarks with a huge grain of salt, but in my own experience, the latest versions of Kimi, MiMo, and GLM (pre-5.2) had already surpassed Sonnet in terms of output quality for a fraction of the price.
With that said, I'm excited to try GLM 5.2 because I still end up reaching for Opus and GPT 5.5 for many tasks because the open models tend to get stuck more often on complex problems.
Do you have benchmarks or at least anecdotes to back that up? I'm not arguing with you; I would just love to see some proof that open models are getting as good as Anthropic's models.
look at benchmarks, use the model yourself. Im usually first to call BS on every chinese model that says they are as good as Opus. this is finally the first one that actually is. It is a massive jump from every other previous chinese model.
I wish I had the time to set it up and work on side projects but unfortunately life and work have been crazy (as I'm sure many here feel). That's why I asked for anecdotes about it.
My Mac Studio uses about 60–80 watts whenever I’m running a model (as measured by the system metrics), so it’s less than 2 kWh/day at full blast. Electricity is like 0.125 €/kWh, so that 24-hour period would be <0.25 €.
Not accounting hardware in my costs, since I didn’t buy my hardware for running models. Running models is just something it can do in addition to what I got it for.
imho everything but opus produces unusable code (fable was even better...), eg gpt5.5 seems to write the absolute worst code that still technically solves the problem; tbh I'd be totally willing to trade "raw intelligence" for "code taste"
more labs need to figure out whatever anthropic did to destroy everybody else on frontiercode bench
Opus has the nickname "Slopus" in a lot of circles for a reason. It can write nice code in isolation, but the way it organizes that code and its rigor in addressing edge cases/making sure things are robust leave a lot to be desired. Opus is particularly famous for having a real problem reinventing stuff that already existed in the codebase because it wanted to get to work before exploring sufficiently.
what you're describing doesn't sound like such a big deal -- it's (A) obvious during review, (B) easy to fix in a single prompt, (C) simple enough to fix manually, (D) can be mitigated with tokenmaxxing (agent review passes, prompting, subagents, etc)
regarding edge cases -- less is more in my experience, as removing is harder than adding
This is a huge blow to Anthropic/OpenAI/Google and a massive win for the rest of the world. The official API prices and speeds mean nothing for open source models.