I like to compare models with a similar score on cost per task and output tokens per task since those measure two things I'm interested in: cost efficiency and token efficiency. Here's how GLM-5.3 compares to other models in a similar score and against GLM-5.2 to save a few clicks for others who care about these metrics:
Model Score Cost / Task Output Tokens / Task
-------------------------------------------------------------------------
GLM-5.3 (max) 59.5 $0.68 41,107
GLM-5.2 (max) 53.0 $0.56 32,200
Claude Opus 5 (high) 61.5 $1.52 21,353
GPT-5.6 Sol (max) 60.9 $1.23 16,879
Grok 4.6 (high) 60.9 $0.84 21,735
Kimi K3 (max) 59.7 $0.84 25,474
GPT-5.6 Sol (xhigh) 59.0 $0.87 11,098
Claude Opus 5 (medium) 58.6 $0.98 12,459
Qwen3.8 Max 58.1 $1.13 38,287
Qwen3.8 2.4T A95B 57.7 $0.95 32,472
Claude Opus 4.8 (max) 57.3 $1.65 33,557
GPT-5.6 Sol (high) 57.3 $0.52 7,545
Muse Spark 1.2 (xhigh) 56.8 $0.40 30,430
GPT-5.6 Terra (max) 56.6 $0.51 20,838
GPT-5.5 (xhigh) 56.3 $0.69 16,893
Gemini 3.7 Flash (high) 56.0 $0.40 36,847
Tested muse spark 1.2 because it was rated so high on design arena, and I've missed a model that can do nice UI in the hands of an operator with no UI skills.
It produced worse UI mockups than GPT and GPT models are already the bottom of the barrel here. The only model that performed well was Kimi K3 - insanely good, but expensive.
If you just want it to generate UI out of nothing, the benchmarks aren't really for that.
If you want to generate a UI based on specific user input of some kind, then they are.
I'd suggest using one model for UI and another model for tacking onto that UI. LLMs are great at pattern matching, and benchmarks don't really capture one-shotting desirable UI.
That said, benchmaxxing is a thing and your experience with models is a thing. Benchmarks are fuzzy and should be taken with a grain of salt.
This matches my experience with Sol. Read and thought for a while, and edited files, tested, edited again, then ran out of budget in a relatively short time. But its solution was very good and was done quickly, so all things equal I prefer that over something much more verbose like Deepseek.
for over 1 billion real world users living in China, they don't have the option of paying $1.52 per task to use Opus 5, they are banned doing that due to US politics.
I've tested GLM 5.3 on the release day and Artificial Analysis is spot on. It's a really good model.
But my main takeaway was something else. I've used closed weight models for long enough that I've forgotten how good it feels to see reasoning tokens.
With GPT/Claude, you kind of hope that intent was captured well, that agent had all the information, all the tools it needed, because you won't see "hmmm it seems like nix flake isn't available here and I shouldn't install something globally" until it slopped out millions of tokens and wasted hundreds of dollars for 8 hours. With GLM and the likes, you just stop the disease right where it begins.
Yes, not necessary often but being able to stop something that is going off the rails is super useful. Especially if the root cause is prompt ambiguity - inject a clarification & it recovers
It's also starting to go beyond reasoning and it's becoming much more problematic. Reasoning is one thing, but codex, for example now encrypts agent-to-agent messages as well, and compaction. I've no idea what subagents are instructed to do, or what they reported back in native codex.
The only thing that's keeping me is the value $200 subscription provides. If that value disappears, I see no reason why not to switch to something that isn't a black box.
I understand that running these benchmarks can get expensive, but it would be really nice to see AA include more benchmarks of models at reasoning settings other than the maximum, at least for the biggest releases. They have that nice graph of cost vs. composite benchmark score with the Pareto frontier line, but who knows if those are actually the optimal choices? There are already a few non-max-reasoning models on the Pareto line, among the few that were tested.
The Chinese models also like to cut corners on stuff like science. Their scores on stuff like biotech and scientific knowledge is far from ChatGPT unfortunately. (Claude is pretty good but it just refuses all prompts).
Very impressive score for the size, though token use is higher than k3 and far higher than proprietary models, and its price to performance isn't all that far ahead of k3 as a result
Is it worth using these models if I have a claude code subscription already? The appeal of lower cost is nice but I haven't gotten over the switching cost yet.
Use a unified proxy that lets you switch between models seamlessly. We are far from an equilibrium in this market and you will continue to have FOMO no matter who you pick if you go all in on one company
I use the $200 plan w/ Anthropic and run out of tokens half way through the week and supposedly they are progressively reducing the limits on all their subs even further.
At some point I will switch, $200 buys a lot of tokens on OpenRouter.
Is the conventional wisdom that the subscription price/token is better than the API price/token not valid any more? Or is access to model diversity worth the increased per token costs?
If anything, it's going to be more expensive. Price/performance ratio isn't there yet for frontier open weight models.
But regardless, you definitely should use a harness where switching models on the fly is easy. There's a reason why Anthropic uses their own proprietary formats/conventions anywhere they can - to lock you in when inference eventually commoditizes.
FYI, you can use your Claude subscription pricing with OpenCode via Meridian[0], which also makes it easier to try out other models when they come out. You can also use your other subscriptions in OpenCode with CLIProxyAPI[1]. The switching cost was relatively high, mostly from claude code plugins but completely worth it. I'm now mostly using GLM-5.3 and Codex models via OpenCode and barely using Claude which seemed unfathomable less than two months ago.
I like to compare models with a similar score on cost per task and output tokens per task since those measure two things I'm interested in: cost efficiency and token efficiency. Here's how GLM-5.3 compares to other models in a similar score and against GLM-5.2 to save a few clicks for others who care about these metrics:
Edited for accuracy and more models.
Muse Spark has a nice balance. not to mentions the Contribs version is old deepseek flash prices.
I found the sweetspot here: GPT-5.6 Sol (high) 57.3 $0.52 7,545
(Edit: TLDR; It gets on with it, makes the same mistakes you would, without overthinking and overengineering, most of the time)
Tested muse spark 1.2 because it was rated so high on design arena, and I've missed a model that can do nice UI in the hands of an operator with no UI skills.
It produced worse UI mockups than GPT and GPT models are already the bottom of the barrel here. The only model that performed well was Kimi K3 - insanely good, but expensive.
It's hard to trust benchmarks these days.
If you just want it to generate UI out of nothing, the benchmarks aren't really for that.
If you want to generate a UI based on specific user input of some kind, then they are.
I'd suggest using one model for UI and another model for tacking onto that UI. LLMs are great at pattern matching, and benchmarks don't really capture one-shotting desirable UI.
That said, benchmaxxing is a thing and your experience with models is a thing. Benchmarks are fuzzy and should be taken with a grain of salt.
these $/task figures aren't very useful in my experience. it doesn't tell you how well it did the task.
generally I choose models by their intelligence and then personal preference from direct experience.
they make a ton of sense for tasks that are achievable with no prob by all models - e.g. writing scripts that do a specific thing etc.
you can use a Gemini model completely for free doing that. no agent. aistudio.google.com
It would make reading and comparing a bit easier if the data was sorted by a dimension.
Cost per task:
Benchmark score:
This matches my experience with Sol. Read and thought for a while, and edited files, tested, edited again, then ran out of budget in a relatively short time. But its solution was very good and was done quickly, so all things equal I prefer that over something much more verbose like Deepseek.
this is not very useful.
for over 1 billion real world users living in China, they don't have the option of paying $1.52 per task to use Opus 5, they are banned doing that due to US politics.
What’s the reseller situation?
I've tested GLM 5.3 on the release day and Artificial Analysis is spot on. It's a really good model.
But my main takeaway was something else. I've used closed weight models for long enough that I've forgotten how good it feels to see reasoning tokens.
With GPT/Claude, you kind of hope that intent was captured well, that agent had all the information, all the tools it needed, because you won't see "hmmm it seems like nix flake isn't available here and I shouldn't install something globally" until it slopped out millions of tokens and wasted hundreds of dollars for 8 hours. With GLM and the likes, you just stop the disease right where it begins.
Yes, not necessary often but being able to stop something that is going off the rails is super useful. Especially if the root cause is prompt ambiguity - inject a clarification & it recovers
It's also starting to go beyond reasoning and it's becoming much more problematic. Reasoning is one thing, but codex, for example now encrypts agent-to-agent messages as well, and compaction. I've no idea what subagents are instructed to do, or what they reported back in native codex.
The only thing that's keeping me is the value $200 subscription provides. If that value disappears, I see no reason why not to switch to something that isn't a black box.
With GPT/Claude, hiding those from users to waste their tokens is a feature, not a limitation.
Still yet, I cannot justify switching from dirt-cheap Luna model, which is pretty damn "intelligent" and works well for my flow
I understand that running these benchmarks can get expensive, but it would be really nice to see AA include more benchmarks of models at reasoning settings other than the maximum, at least for the biggest releases. They have that nice graph of cost vs. composite benchmark score with the Pareto frontier line, but who knows if those are actually the optimal choices? There are already a few non-max-reasoning models on the Pareto line, among the few that were tested.
You can turn on various levels of some of many of the models in the UI
Beware of the benchmarks listed. SciCode and EnterpriseOps for instance: https://shukla.io/blog/2026-08/gym.html
The Chinese models also like to cut corners on stuff like science. Their scores on stuff like biotech and scientific knowledge is far from ChatGPT unfortunately. (Claude is pretty good but it just refuses all prompts).
Does Artificial Analysis use OpenRouter for model access to do their benchmarks?
Sol is an underappreciated model. Dropped Claude today and went to codex. None of that god awful prose Claude used for me any longer.
Very impressive score for the size, though token use is higher than k3 and far higher than proprietary models, and its price to performance isn't all that far ahead of k3 as a result
>token use is higher than k3 and far higher than proprietary models
GLM sets effort to max by default historically.
Aa also benchmarked k3 at max
Is it worth using these models if I have a claude code subscription already? The appeal of lower cost is nice but I haven't gotten over the switching cost yet.
At least by API usage, they aren't yet lower cost than subscriptions. Not sure about GLM's subscription plans though.
GLM subscription is better than API, but significantly worse than Codex, even when used outside peak hours.
Use a unified proxy that lets you switch between models seamlessly. We are far from an equilibrium in this market and you will continue to have FOMO no matter who you pick if you go all in on one company
no, at subscription prices claude is a better value than GLM.
They're only a better value if you're paying API rates
I use the $200 plan w/ Anthropic and run out of tokens half way through the week and supposedly they are progressively reducing the limits on all their subs even further.
At some point I will switch, $200 buys a lot of tokens on OpenRouter.
Same here. I'm switching to the Codex plan. It just doesn't go very far now. Especially if you use fable at all.
Is the conventional wisdom that the subscription price/token is better than the API price/token not valid any more? Or is access to model diversity worth the increased per token costs?
If anything, it's going to be more expensive. Price/performance ratio isn't there yet for frontier open weight models.
But regardless, you definitely should use a harness where switching models on the fly is easy. There's a reason why Anthropic uses their own proprietary formats/conventions anywhere they can - to lock you in when inference eventually commoditizes.
FYI, you can use your Claude subscription pricing with OpenCode via Meridian[0], which also makes it easier to try out other models when they come out. You can also use your other subscriptions in OpenCode with CLIProxyAPI[1]. The switching cost was relatively high, mostly from claude code plugins but completely worth it. I'm now mostly using GLM-5.3 and Codex models via OpenCode and barely using Claude which seemed unfathomable less than two months ago.
[0] https://github.com/rynfar/meridian
[1] https://github.com/router-for-me/CLIProxyAPI
edit: reworded for clarity
Tied for #1 by agentic index (with Opus 5).
And reminder: it's less than a quarter the size of Kimi K3!
...do I take out a double mortgage to buy a 4 Spark cluster?
$20k is personal loan territory, not a second mortgage lol
Not with my credit!
No, you use openrouter and spend 10% as much as using a proprietary model.
Qwen3.8 27B doing a lot of lifting right now, and people seem to run it pretty well on 1-2x 3090 setups...