- 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.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- Muse Spark has a nice balance. not to mentions the Contribs version is old deepseek flash prices.
- 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.
- 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)
- It would make reading and comparing a bit easier if the data was sorted by a dimension.
- Cost per task:
Benchmark score:Model Score Cost / Task Output Tokens / Task ------------------------------------------------------------------------- Muse Spark 1.2 (xhigh) 56.8 $0.40 30,430 Gemini 3.7 Flash (high) 56.0 $0.40 36,847 GPT-5.6 Terra (max) 56.6 $0.51 20,838 GPT-5.6 Sol (high) 57.3 $0.52 7,545 GLM-5.2 (max) 53.0 $0.56 32,200 GLM-5.3 (max) 59.5 $0.68 41,107 GPT-5.5 (xhigh) 56.3 $0.69 16,893 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 Qwen3.8 2.4T A95B 57.7 $0.95 32,472 Claude Opus 5 (medium) 58.6 $0.98 12,459 Qwen3.8 Max 58.1 $1.13 38,287 GPT-5.6 Sol (max) 60.9 $1.23 16,879 Claude Opus 5 (high) 61.5 $1.52 21,353 Claude Opus 4.8 (max) 57.3 $1.65 33,557Model Score Cost / Task Output Tokens / Task ------------------------------------------------------------------------- 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 GLM-5.3 (max) 59.5 $0.68 41,107 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 GLM-5.2 (max) 53.0 $0.56 32,200- 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.
- 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.
- 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.
- Sol is an underappreciated model. Dropped Claude today and went to codex. None of that god awful prose Claude used for me any longer.
- To me one of the biggest limitations of GLM is the lack of multi-modality.
For web dev is just a must to have, and offloading that part to a secondary model doesn't work really well in my experience.
- 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?
- 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
- Yes, they have multiple levels of Claude, GPT, Gemini, and Kimi, but not the other top models (I would put GLM, Qwen, Muse, Grok, and Deepseek in that bucket).
- 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.
- 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
- 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
- Yes. Please seriously try other models. See relevant thread here: https://news.ycombinator.com/item?id=49296740
- 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.
- 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.
- 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.
- Tied for #1 by agentic index (with Opus 5).
- And reminder: it's less than a quarter the size of Kimi K3!
- This comment by the Z AI lead is relevant, about parameter vs data scaling: https://news.ycombinator.com/item?id=49357405
- ...do I take out a double mortgage to buy a 4 Spark cluster?
- 20k is credit card territory
- 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...
- $20k is personal loan territory, not a second mortgage lol
- Not with my credit!
- [flagged]
- [flagged]
