One thing about these numbers that's absolutely shocking to me is how low the energy use is:
> That model’s usage was well within our budget ($68, about 4kWh of energy use / 365 grams of carbon emissions).
The energy cost is literally 1% of the total cost. For context, 4kWh of energy would drive you about 15 miles in an EV, about half of the average person's daily driving miles. It's boiling 10 gallons of water.
With the talk of AI data centers' impact on the world, you'd think this would be 10x to 100x the amount of energy in order to get the effects they're using here.
My takeaway: the AI data center buildout is an overbuild probably at least as large as the fiber buildout that left us with so much dark fiber. If not even bigger. The only thing that will save the economy is the inability of NVIDIA and chip fabs to produce enough chips to match the buildout planned.
I have the feeling China is somehow ahead when it comes to energy (and cost) efficiency for AI usage. After all, the two are in a direct competition, and this difference is significant. Or is the "hyper" scaling of energy hungry datacenters in US part of a bubble?
Your only talking about variable direct energy. Does it take into account the entire lifecycle, building the data centre, running the cooling, building the chips, the % the chips are not utilised.
I agree but it does worry me how fast my usage is increasing. Two months ago I was using 10x less tokens and probably not much more than 5kWh on inference. This month about 30kWh on inference. If it becomes more affordable, is there going to be another jump? Not quite sure
At some point the agents will be good enough that you can tell them "here's $100, go make me money" and they will, maybe not a lot and not all the time, but the EV given the cost of inference will be positive.
> Unfortunately there are still consequences to it. I chose the 'wrong' model for the prototype, and we spent 450M tokens / $150 / 5kWh of energy use almost overnight. The MCP server itself works well and we now have a great demo of the capabilities, so it’s not for nothing:
> Nonetheless, it’s a good reminder to be careful with model selection and with agentic patterns. We could have achieved similar results for most likely 5x less cost with not that much more effort. Lessons learned! We need to budget for this, and be more careful. Could have seen it coming, but now we know.
I don't get it. Why was it wrong? Which one would have been better? What was the lesson and how could you have foreseen it?
Hmm I might need to rephrase. Initial challenge was to use GLM 5.3 Flash and I was on the non-Flash version for the whole vibe coded build. Just wasn’t paying attention and I didn’t realize that one session was a quarter of the month’s spend and 150% of the budget (big price difference between models)
It was a bit of a silly challenge, wasn’t sure how workable, learned a lot in the process about what actually drives usage / costs, and how to keep both under control
Not OP, but I've used both quite a bit and I think GLM 5.3 (non-flash) is a vastly better model. The flash variant is good for general workhorse agents, but it doesn't seem to reason holistically about code and over-engineers solutions to each specific problem it solves. But if you use GLM 5.3 to write a detailed plan with little to no ambiguity, GLM 5.3 Flash executes it just fine for a fraction of the price.
I have a hard time justifying GLM 5.3 these days. It’s slightly better than Flash but rarely enough to justify the much steeper price. We chose to use usage-based billing only so are very sensitive to model price.
When text wuality or for pure but adwansed coding is concerned i always pick glm 5.3. The flash is awesome for everything that dosent really matter though.
Ill tell you from my personal use glm 5.3 flash was better then deepseek 4.1 flash, also deepseek liked to yap in his reasoning traces soo fucking much, the yapping was fast but the task was so slow to complete...
Interesting, I have roughly the opposite experience regarding speed: DeepSeek V4.1 Flash gets stuff done way more quickly for me than GLM 5.3 Flash. (I was getting 300+ tokens/sec with DS vs ~100 with GLM in my testing.) I agree that GLM 5.3 Flash is a slightly better model, but for me at least, it's not a huge difference, and I'd rather have DeepSeek's speed.
Flash is pretty decent coder, but it should be paired with good planner and reviewer. I would pick astra low for planning and sol 6.1 medium for reviews.
It's my favorite model family to interact with, it's prose is the best imo, it makes me laugh from time-to-time (like when it said it would "crib" some code from another project, lul)
I currently have qwen-flash working on an NES emulator harness so qwen-little can play my first RPG (ff1)
(tho I have used all the others I mentioned, happenstance I'm using qwen this iteration/task)
There are some issue point not properly mentioned ...
* Models like DeepSeek V4.1 Flash are much cheaper on DeepSeek their API directly because of the cache handeling is better. Neuralwatt can hit up to 98% but DeepSeek can do 99.x... That may not sound like a big difference but it quickly widens the gap on long tasks to grow 2x a 3x in price. DeepSeek their cache handeling is S-tier (with a ton of features, for instance 24h caching).
* The same issue is also present if you compare GLM 5.3 Flash with z.ai vs Neuralwatt. Its just way more cheaper from the source, then from Neuralwatt.
* The energy numbers from Neuralwatt are ... to be taken with a ton of salt. Past energy numbers had the same models (for instance) GLM 5.2 up to 6x cheaper in energy usage, then after they "fixed" issues with the energy numbers. In reality, those energy numbers are just a different form of billing, but not a actual representation of the energy usage of AI models. Things like profits are inside those energy numbers. So seeing 4kWH used for a model, does not mean that it uses 4Kwh.
For what it’s worth the energy numbers I get from Neuralwatt are within the ballpark of what is available elsewhere like https://cleerdash.sustainableaigroup.com/. What’s your source that there is profit / capex in there? Their published methodology seems pretty transparent
I'm skeptical that the DeepSeek official API actually nets out cheaper right now, but I avoid it anyway because they store and train off of your prompts.
I think NW's profits are mostly between what they pay for electricity and what they charge you for electricity. I don't think there's any need to look to conspiracies to explain billing errors.
One thing about these numbers that's absolutely shocking to me is how low the energy use is:
> That model’s usage was well within our budget ($68, about 4kWh of energy use / 365 grams of carbon emissions).
The energy cost is literally 1% of the total cost. For context, 4kWh of energy would drive you about 15 miles in an EV, about half of the average person's daily driving miles. It's boiling 10 gallons of water.
With the talk of AI data centers' impact on the world, you'd think this would be 10x to 100x the amount of energy in order to get the effects they're using here.
My takeaway: the AI data center buildout is an overbuild probably at least as large as the fiber buildout that left us with so much dark fiber. If not even bigger. The only thing that will save the economy is the inability of NVIDIA and chip fabs to produce enough chips to match the buildout planned.
Related to this: see the difference between China and US here: https://aidatacenterindex.com/compare/united-states-vs-china...
I have the feeling China is somehow ahead when it comes to energy (and cost) efficiency for AI usage. After all, the two are in a direct competition, and this difference is significant. Or is the "hyper" scaling of energy hungry datacenters in US part of a bubble?
It would be a pretty big deal if the average person's car became 50% less energy efficient, no?
Your only talking about variable direct energy. Does it take into account the entire lifecycle, building the data centre, running the cooling, building the chips, the % the chips are not utilised.
No those numbers are GPU only. See https://cleerdash.sustainableaigroup.com/ for a fuller model
I agree but it does worry me how fast my usage is increasing. Two months ago I was using 10x less tokens and probably not much more than 5kWh on inference. This month about 30kWh on inference. If it becomes more affordable, is there going to be another jump? Not quite sure
At some point the agents will be good enough that you can tell them "here's $100, go make me money" and they will, maybe not a lot and not all the time, but the EV given the cost of inference will be positive.
> Unfortunately there are still consequences to it. I chose the 'wrong' model for the prototype, and we spent 450M tokens / $150 / 5kWh of energy use almost overnight. The MCP server itself works well and we now have a great demo of the capabilities, so it’s not for nothing:
> Nonetheless, it’s a good reminder to be careful with model selection and with agentic patterns. We could have achieved similar results for most likely 5x less cost with not that much more effort. Lessons learned! We need to budget for this, and be more careful. Could have seen it coming, but now we know.
I don't get it. Why was it wrong? Which one would have been better? What was the lesson and how could you have foreseen it?
Hmm I might need to rephrase. Initial challenge was to use GLM 5.3 Flash and I was on the non-Flash version for the whole vibe coded build. Just wasn’t paying attention and I didn’t realize that one session was a quarter of the month’s spend and 150% of the budget (big price difference between models)
GLM5.3-flash has been fantastic for me to make minor fixes in ambigious ways. "Fix x feature, whats going wrong. " It does the job.
GLM-5.3-flash is my implementation model after GLM-5.3 writes the plan.
It's an excellent workhorse. When I am running out of my GLM quota I switch GLM-5.3-flash to DS-4.1-flash.
Was this post generated with LLM, did he properly mention anywhere why exactly did it fail with example or i have trouble reading.
It was a bit of a silly challenge, wasn’t sure how workable, learned a lot in the process about what actually drives usage / costs, and how to keep both under control
Considering they were your top two models, how did the flash and non-flash versions compare? Did you use them for different tasks?
Not OP, but I've used both quite a bit and I think GLM 5.3 (non-flash) is a vastly better model. The flash variant is good for general workhorse agents, but it doesn't seem to reason holistically about code and over-engineers solutions to each specific problem it solves. But if you use GLM 5.3 to write a detailed plan with little to no ambiguity, GLM 5.3 Flash executes it just fine for a fraction of the price.
I have a hard time justifying GLM 5.3 these days. It’s slightly better than Flash but rarely enough to justify the much steeper price. We chose to use usage-based billing only so are very sensitive to model price.
When text wuality or for pure but adwansed coding is concerned i always pick glm 5.3. The flash is awesome for everything that dosent really matter though.
Worth a comparison with DeepSeek v4.1 flash, if you've got another month to spare!
Ill tell you from my personal use glm 5.3 flash was better then deepseek 4.1 flash, also deepseek liked to yap in his reasoning traces soo fucking much, the yapping was fast but the task was so slow to complete...
Interesting, I have roughly the opposite experience regarding speed: DeepSeek V4.1 Flash gets stuff done way more quickly for me than GLM 5.3 Flash. (I was getting 300+ tokens/sec with DS vs ~100 with GLM in my testing.) I agree that GLM 5.3 Flash is a slightly better model, but for me at least, it's not a huge difference, and I'd rather have DeepSeek's speed.
Yep, I think next month will be on that. It feels slightly better from a few days of use, and in our WIP benchmarking it scores way higher
DS v4.1 Flash is roughly equivalent. It's going to get some things right/better that GLM flash doesnt and vice versa.
They've got an image of their homegrown benchmark in the post that lists DS4.1.
How did you measure energy usage?
Edit: I found a linked article that mentions the inference provider who does the measurements.
Yes, all from Neuralwatt, GPU energy use only. Makes models’ “efficiency” much more visible than tokens.
Flash is pretty decent coder, but it should be paired with good planner and reviewer. I would pick astra low for planning and sol 6.1 medium for reviews.
What would you use if you wanted to stay (at least) open weight?
qwen3.8, kimi3, kimi2.7, GLM-5.3 are all good families I use in my coding team
I'm mainly using flash varients, at least as the default, bump.up to stronger model as needed (less often these days)
qwen3.8 27B or 2.4T? they are completely different models with completely different pricing
I use all the qwen!
It's my favorite model family to interact with, it's prose is the best imo, it makes me laugh from time-to-time (like when it said it would "crib" some code from another project, lul)
I currently have qwen-flash working on an NES emulator harness so qwen-little can play my first RPG (ff1)
(tho I have used all the others I mentioned, happenstance I'm using qwen this iteration/task)
Deepseek 4.1 Flash and Mimo 2.6 Flash.
Flash is plenty good for planning and reviewing, for my needs. In fact, I use it for that because it's too slow for execution, despite the name.
edit: I subscribe to z.ai, I don't host.
> I use it for that because it's too slow for execution
What kind of hardware and what particular quant?
Will agree on this. From z.ai i have found the non flash to have way more consistent performance.
There are some issue point not properly mentioned ...
* Models like DeepSeek V4.1 Flash are much cheaper on DeepSeek their API directly because of the cache handeling is better. Neuralwatt can hit up to 98% but DeepSeek can do 99.x... That may not sound like a big difference but it quickly widens the gap on long tasks to grow 2x a 3x in price. DeepSeek their cache handeling is S-tier (with a ton of features, for instance 24h caching).
* The same issue is also present if you compare GLM 5.3 Flash with z.ai vs Neuralwatt. Its just way more cheaper from the source, then from Neuralwatt.
* The energy numbers from Neuralwatt are ... to be taken with a ton of salt. Past energy numbers had the same models (for instance) GLM 5.2 up to 6x cheaper in energy usage, then after they "fixed" issues with the energy numbers. In reality, those energy numbers are just a different form of billing, but not a actual representation of the energy usage of AI models. Things like profits are inside those energy numbers. So seeing 4kWH used for a model, does not mean that it uses 4Kwh.
For what it’s worth the energy numbers I get from Neuralwatt are within the ballpark of what is available elsewhere like https://cleerdash.sustainableaigroup.com/. What’s your source that there is profit / capex in there? Their published methodology seems pretty transparent
I'm skeptical that the DeepSeek official API actually nets out cheaper right now, but I avoid it anyway because they store and train off of your prompts.
I think NW's profits are mostly between what they pay for electricity and what they charge you for electricity. I don't think there's any need to look to conspiracies to explain billing errors.