Using GLM 5.2 with Neuralwatt's Energy-Based Billing
Hello, I'm Munou.
I've been wanting to use GLM 5.2 recently but have been struggling to figure it out, so I'm writing this down.
Research
First, I researched where it could be used for free.
Cloudflare Workers AI offers a free tier for now, so I tried it. Wow, it's way too slow.
I guess that's to be expected, but perhaps because of the free tier, the response speed is terrible.
https://openrouter.ai/z-ai/glm-5.2
Currently, it takes 43 seconds just for the API response, so it's unusable. In fact, when I tried the official Z.ai site that I was using for testing, I didn't get any response at all, which is pretty bad.
It seems to be offered in Open Code's subscription plan as well, but the reviews from Redditors aren't very good, and the token consumption is so high that you hit the limit quickly, making it difficult to manage.
Also, if a model becomes less intelligent when compressed through quantization, there's no point in paying for it if it ends up being unusable. So, I decided to just scour the internet for reviews and interesting options. I registered with a few Chinese providers, but I stopped because information about the models being used wasn't made public.
Ultimately, Hacker News users were recommending a place called Neuralwatt, so I looked into it and found they have an interesting pricing model called energy-based billing.
The cost of use is variable depending on this power consumption, and you can check it on the official page as shown below.
https://portal.neuralwatt.com/pricing
You can also check the energy status to see how much power was consumed in a single request.
https://portal.neuralwatt.com/energy-pricing
Comparison with Codex
Based on my experience, GLM-5.2 (short, fast) has a response speed equivalent to Codex, which I use as my main tool, but in this case, the context is 200K.
The GPT 5.5 Medium used in Codex includes reasoning, so the maximum context length is likely 256k for combined input/output. To treat them as effectively equivalent, GLM-5.2 (short) would be the choice. Note that the "Fast" model lacks reasoning.
OpenAI GPT 5.5 :
https://help.openai.com/ja-jp/articles/11909943-gpt-55-in-chatgpt
Neuralwatt GLM-5.2 (short, fast) :
https://portal.neuralwatt.com/models/glm-5.2-short-fast
Neuralwatt GLM-5.2 (short) :
I originally switched to Codex because the response from Gemini CLI was too slow, and I was so impressed that I subscribed. In my experience, GLM-5.2 (short) is practically at the same speed level as Codex 5.5 Medium, making it very practical.
The ability to easily switch models when reasoning isn't even needed is a benefit of using Open Code. Regarding bug investigation, there were instances where a bug not found by Codex GPT 5.5 xhigh with the same prompt was discovered by the GLM 5.2 1048K context model. In this regard, I feel a slight difference in the training models, but for the most part, they are about the same.
If they are about the same, I sometimes think it might be more economical to use a pay-as-you-go system where the balance doesn't decrease much. However, the cost-performance of the Codex quota available through a ChatGPT subscription is hard to give up, so it's a tough call.
However, my gut feeling is that if Codex becomes temporarily unavailable, or if you've used up your quota and don't want to add more subscriptions, or if Codex quota consumption accelerates, it's a very strong candidate as a backup LLM.
After using GLM 5.2 with a 1048K maximum context quite a bit during my off-hours, my balance went from an $11 charge to $8.09 remaining over 24 hours, so it cost about $3.
Considering it can handle bug investigations and perspective corrections that GPT 5.5 misses, being able to use it at this price point seems like a great option for Codex users who don't want to deal with Claude (which requires a subscription and has no free tier).
Other Models
I think Xiaomi's MiMo 2.5 is actually quite useful, but there's still so little information that I haven't been able to research it fully yet.
https://mimo.xiaomi.com/mimo-v2-5/
It's available for free on Open Code Go, and it feels like there are no limits even if you keep running it, so being able to use it essentially unlimited is quite good, but since I don't know the model details of MiMo 2.5 Free itself, I can't say for sure. It's excellent enough that I'm considering paying a bit to use it. Especially since, even though it's likely used for training, if you leave it for a while, it can handle a fair amount of corrections and app creation, so I feel like I can leave tasks like creating throwaway scripts to it.