Several factual errors about the model here. The input modalities are listed as text only, but the headline feature is image support. The context length should be 1048576 (so should GLM-5.3's, also wrong on the charts).
I have both GLM & Gemini in a subscription and see no reason for choosing GLM 5.3 Flash. Working with de speed of Gemini 3.7 Flash is such a delight that I accept the hassle of working with Antigravity CLI, coming from Claude Code which I use for GLM.
I find the Time per Task[0] metric more helpful, because models vary enormously in the tokens required to complete a task. On Time per Task, Gemini 3.7 Flash is Matched with GPT-5.6-Sol, as well as on price per task.
GLM-5.3-Flash takes 7x (relative to Gemini and Sol) per task. So, it's cheaper, if you don't value your time! Don't value real-time workflows, don't value iteration speed, etc. So, doesn't seem very suitable for interactive or agentic work to me.
But having an ultra cheap model for async stuff is always very nice. (Still, the last few weeks feel less about tech and more like a contest between who can afford to give the biggest discounts!)
--
I also like DeepSwe[2], although they measure Output Tokens and Agent Steps, which are misleading when one model has a much faster output speed. (e.g. on their metrics Gemini looks slower, because they don't account for that.)
Been using Luna exclusively since the price drop, and i've been very satified with all tasks from planning, writing code, and other agent tasks. (just change thinking level from low <-> ultra)
---
btw, I did try out Ox Alpha, the coding feels good but still not way better for me to switch to it.
Luna is at a very compelling point on the price/performance curve.
I have found that sometimes a smaller model with max reasoning is actually more expensive than using the next tier model with a lower reasoning effort. It’s certainly faster.
I’ve been super pro-Luna lately. I really hope that Gemini-Flash-Lite is positioned to compete with it. We all know that Anthropic has abandoned Haiku and it would never be that cheap.
Probably shouldn’t say this here but I’ve been planning to up my $20/mo exploratory ChatGPT subscription to the $100/mo tier as soon as I hit my cap. Between the progress and quality of Luna and their continuous resets, it’s been a few months now that I’ve lived off the $20 tier, frankly waiting for the need to upgrade, credit card in hand.
I’m always trying new models, like many of us here, but the price is just so good for a well balanced, American, hosted model.
I don't understand what would possibly make someone prefer speed over output? You'd rather get wrong bad answers that don't work as well very fast?
In general I really don't mind waiting 5, 10, 40 minutes. There's other things I can look at, other plans or assessments or outputs aplenty stacking up. Its baffling beyond words to me that anyone would take speed over good output. Surely the better output is going to save enormous time in the long run, have better outcomes. What is it that addicts people so much to speed, especially when the difference is between fast and very fast?
Background agent services that require no super human vision.
You might think faster is better when you vibe code or monitoring.
But with background agents, the faster the speed, the more jobs it can perform.
The speed won't matter as much for your personal projects, but if you want to handle enterprise level request, the faster the better.
Say you queued up all messages for a task in say Kafka, you got workers calling AI agents. You will have thousands of messages to do, and the faster the AI agents can do its work the better you can clear the queue.
> So how exactly is Anthropic and OpenAI ever going to pay back the trillions that they plan on spending?
It's really simple: if they truly get to human-level AI (or even superhuman AI), then money and debts no longer matter, since our current economic system will be obsolete. They are betting everything on this outcome.
I don't know if they will manage to do it before their debts have to be repaid, but considering the rate of acceleration in the past few months, there is a non-trivial chance that they will, IMHO. We will see.
1. they still have revenue though. it might not enough to cover all the r&d but it is surely enough to cover the hardware cost.
2. people tend to ignore this, but the salary budget of a US frontier lab and chinese frontier lab is nowhere comparable, the first can easily outdone the later by 100x.
3. us labs, like other US style startups, always throw ton of money to capture the market. I don't see the chinese company doing the same scheme at all.
so, surely chinese AI providers also lost money making new models, but they are not spending nearly as much as US ones.
By charging $$$ like they do now and having a non terminal business model.
PRC AI have lower opex and capex, i.e. export controls means they couldn't be trillions in the hole on inflated hardware in the first place. They only need to extract a few 10s of billions from domestic market have a healthy runway. If investors/gov wants to throw in a few billion to treat as utility, whatever, it's still rounding error.
just like there were mistrals, coheres, llamas, etc, there will be new deepseeks and moonshots if those ever flame out (worst case, given out at cost by google, meta, alibaba or etc)
OpenAI and Anthropic are already in a ~200bil hole from previous model iterations and are committing to trillions of additional spending
OpenAI spent more TBPN than kimi spent on training K3
By charging $$$ like they do now and having a non terminal business model. PRC AI have lower opex and capex, i.e. export controls means they couldn't be trillions in the hole on inflated hardware in the first place. They only need to extract a few 10s of billions from domestic market have a healthy runway. If investors/gov wants to throw in a few billion to treat as utility, whatever, it's still rounding error.
I expect they're going to fight each other to become the vendor of record for the government, and whoever wins will get bailed out. This is one area where they don't have to worry about competition from Chinese models.
Its more Google Amazon Meta Microsoft who are spending trillions. They will be fine. So will Anthropic and OpenAI. Nvidia will presumably survive. The losses are all the real estate interests and contractors and contributory hardware companies etc.
Several factual errors about the model here. The input modalities are listed as text only, but the headline feature is image support. The context length should be 1048576 (so should GLM-5.3's, also wrong on the charts).
https://docs.z.ai/guides/vlm/glm-5.3-flash#model-api
Looking at these numbers IMHO, with Gemini you get the speed what you pay for.
I have both GLM & Gemini in a subscription and see no reason for choosing GLM 5.3 Flash. Working with de speed of Gemini 3.7 Flash is such a delight that I accept the hassle of working with Antigravity CLI, coming from Claude Code which I use for GLM.
I find the Time per Task[0] metric more helpful, because models vary enormously in the tokens required to complete a task. On Time per Task, Gemini 3.7 Flash is Matched with GPT-5.6-Sol, as well as on price per task.
GLM-5.3-Flash takes 7x (relative to Gemini and Sol) per task. So, it's cheaper, if you don't value your time! Don't value real-time workflows, don't value iteration speed, etc. So, doesn't seem very suitable for interactive or agentic work to me.
But having an ultra cheap model for async stuff is always very nice. (Still, the last few weeks feel less about tech and more like a contest between who can afford to give the biggest discounts!)
--
I also like DeepSwe[2], although they measure Output Tokens and Agent Steps, which are misleading when one model has a much faster output speed. (e.g. on their metrics Gemini looks slower, because they don't account for that.)
[0] Time per Task - https://artificialanalysis.ai/?models=glm-5-3-flash%2Cgemini...
[1] Output Tokens Per Task - https://artificialanalysis.ai/?models=glm-5-3-flash%2Cgemini...
[2] https://deepswe.datacurve.ai/
The analysis is still not compelling for me to switch from gtp5.6-luna to GLM-5.3-flash given
- costs per task $0.05 vs $0.09
- speed 130 vs 88
- where GLM has only 5 more intelligence point: at this point few point is meaningless for most of models
https://artificialanalysis.ai/models/comparisons/glm-5-3-fla...
Been using Luna exclusively since the price drop, and i've been very satified with all tasks from planning, writing code, and other agent tasks. (just change thinking level from low <-> ultra)
---
btw, I did try out Ox Alpha, the coding feels good but still not way better for me to switch to it.
Luna is at a very compelling point on the price/performance curve.
I have found that sometimes a smaller model with max reasoning is actually more expensive than using the next tier model with a lower reasoning effort. It’s certainly faster.
Agreed. with "Ultra" (higher than Max), the luna performs really well for my non-metric-backed personal experience
Which has a better monthly plan? Right now Z.ai "Pro" plan (the middle one) is $56/mo if you prepay for a year.
I signed up for their Lite plan when it was only $28 for the whole year (less than $3/mo). Definitely very happy with that purchase!
I’ve been super pro-Luna lately. I really hope that Gemini-Flash-Lite is positioned to compete with it. We all know that Anthropic has abandoned Haiku and it would never be that cheap.
Probably shouldn’t say this here but I’ve been planning to up my $20/mo exploratory ChatGPT subscription to the $100/mo tier as soon as I hit my cap. Between the progress and quality of Luna and their continuous resets, it’s been a few months now that I’ve lived off the $20 tier, frankly waiting for the need to upgrade, credit card in hand.
I’m always trying new models, like many of us here, but the price is just so good for a well balanced, American, hosted model.
Funny, I used to use Gemini before Luna as it was "good enough" and cheap.
For me at this point, most of newer models are capable enough, I focus more on $ and how much I can save.
> The analysis is still not compelling for me to switch from gtp5.6-luna to GLM-5.3-flash given ...
So Luna is competitive because a few weeks ago they did a 80% price drop?
Many here said that 80% drop was not a move against Anthropic but a move against chinese models and your comments indicate that's the case.
I didn't bat an eye before 80% price drop. I used GLM-5.2 and Gemini flash
I don't understand what would possibly make someone prefer speed over output? You'd rather get wrong bad answers that don't work as well very fast?
In general I really don't mind waiting 5, 10, 40 minutes. There's other things I can look at, other plans or assessments or outputs aplenty stacking up. Its baffling beyond words to me that anyone would take speed over good output. Surely the better output is going to save enormous time in the long run, have better outcomes. What is it that addicts people so much to speed, especially when the difference is between fast and very fast?
Background agent services that require no super human vision.
You might think faster is better when you vibe code or monitoring. But with background agents, the faster the speed, the more jobs it can perform.
The speed won't matter as much for your personal projects, but if you want to handle enterprise level request, the faster the better.
Say you queued up all messages for a task in say Kafka, you got workers calling AI agents. You will have thousands of messages to do, and the faster the AI agents can do its work the better you can clear the queue.
So how exactly is Anthropic and OpenAI ever going to pay back the trillions that they plan on spending?
> So how exactly is Anthropic and OpenAI ever going to pay back the trillions that they plan on spending?
It's really simple: if they truly get to human-level AI (or even superhuman AI), then money and debts no longer matter, since our current economic system will be obsolete. They are betting everything on this outcome.
I don't know if they will manage to do it before their debts have to be repaid, but considering the rate of acceleration in the past few months, there is a non-trivial chance that they will, IMHO. We will see.
That’s like building F1 cars and thinking you will soon have a rocket to land on the moon.
LLM has nothing to do with AGI.
5 replies →
That's assuming that human-level AI is a possibility with current approaches.
3 replies →
You could ask the same about how z.ai, moonshot ai, minimax, and alibaba are going to continue training and releasing models for free.
1. they still have revenue though. it might not enough to cover all the r&d but it is surely enough to cover the hardware cost.
2. people tend to ignore this, but the salary budget of a US frontier lab and chinese frontier lab is nowhere comparable, the first can easily outdone the later by 100x.
3. us labs, like other US style startups, always throw ton of money to capture the market. I don't see the chinese company doing the same scheme at all.
so, surely chinese AI providers also lost money making new models, but they are not spending nearly as much as US ones.
3 replies →
By charging $$$ like they do now and having a non terminal business model.
PRC AI have lower opex and capex, i.e. export controls means they couldn't be trillions in the hole on inflated hardware in the first place. They only need to extract a few 10s of billions from domestic market have a healthy runway. If investors/gov wants to throw in a few billion to treat as utility, whatever, it's still rounding error.
2 replies →
just like there were mistrals, coheres, llamas, etc, there will be new deepseeks and moonshots if those ever flame out (worst case, given out at cost by google, meta, alibaba or etc)
OpenAI and Anthropic are already in a ~200bil hole from previous model iterations and are committing to trillions of additional spending
OpenAI spent more TBPN than kimi spent on training K3
They are owned by the state, so the economics are a bit different.
6 replies →
By charging $$$ like they do now and having a non terminal business model. PRC AI have lower opex and capex, i.e. export controls means they couldn't be trillions in the hole on inflated hardware in the first place. They only need to extract a few 10s of billions from domestic market have a healthy runway. If investors/gov wants to throw in a few billion to treat as utility, whatever, it's still rounding error.
At a fraction of the cost.
2 replies →
Socialize the lost.They dont have to. You pay.
I expect they're going to fight each other to become the vendor of record for the government, and whoever wins will get bailed out. This is one area where they don't have to worry about competition from Chinese models.
Its more Google Amazon Meta Microsoft who are spending trillions. They will be fine. So will Anthropic and OpenAI. Nvidia will presumably survive. The losses are all the real estate interests and contractors and contributory hardware companies etc.
Why do you think tech oligarchs have been cozying up to the Trump admin? They're angling for a government bailout, paid for by your tax money!
Thy cozy up to whoever is in government. They cozy'd up with Biden too.
2 replies →
Impressive. It kicked everything between itself and Sol xhigh out of the Pareto frontier. Can't wait to try it out.
Can’t wait to try this out, and the only missing from this model for me is Image input support
It has image/video input support (that is surprisingly good)
Better than the latest Deepseek v4 Pro while being 3x cheaper in cost per task. Impressive!
Why does the top card say "Intelligence #1/173" when the bar chart further down shows it only at position 7?
And the model isn't even shown in the speed bar chart just below. Such slop (the artificial intelligence website linked)
Not clear to me how it selects that, but it considers it to be #1 out of 173 models "of the same class".
I see, thanks, then it should state that in the card.
[dead]