I just tested GLM 4.5 and 4.5-AIR, the performance for the price is absolutely insane! I really think this should be the default model for auto mode, as mentioned earlier in the conversation. I can’t wait to have access to it on Cursor!
2 Likes
I agree with you, I’m phrasing my request like this on purpose because it’s clear to me the cursor team will not listen to just “normal” requests, a 15-minute call is an incredibly low entry and barrier, and I have provided millions of dollars in improvements to other compaines operations so im not speaking from nowhere, I have legitamate and well thought out ideas and questions sitting on my desk ready for if the cursor team wants to hear them.
Cursor is acting like they are the only business that has ever faced these issues, other businesses have fixed these issues in the past, so i dont even need to making anything new up.
Also cursor has already directly took ideas i have put on these forums and speedily put them into cursor, they were extremely easy to implement, and I had thought about them months ago.
So yes I understand why you would be skeptical of a user claiming they could save millions for a company simply in 15 min conversation, however I have spent hours thinking these solutions through, so all i need is a 15 min call to explain exactly how they can implement it.
2 Likes
In my experience this model is around sonnet level, while having better design patterns in code. I’ve used it since release on various tech like frontend (react) and go backend, as well as a little bit of python scripting. It somehow always nail every task I give it. Would love to see it in cursor!
2 Likes
Am I the only person here who thinks that GLM 4.5 is not even close to Sonnet 4 in terms of code quality? 
1 Like
I tried it with claude code on a flutter project, it did some okay edits but was not able to make the file compile (500 lignes) !
Curious to see in cursor or in other languages, but nothing like Claude 4 sonnet for flutter/dart
2 Likes
i agree, same with qwen3, it is likely because they lack the training on actual knowledge and are mostly trained on ai responses (“distillation”), based on what i understand at least. this makes them cheap and in some instances very fast, which is why i still want them in cursor.
i can run a fast test with qwen3, have it finished in 2-4 seconds, try the result, if it does not work retry a couple of times and then if none of the attempts worked, i can switch to claude. this way i can check if i can get working demo within seconds for just couple of cents before wasting minutes and actual dollars.
1 Like
I think it depends on the tech stack you’re working on and what kind of stuff you’re building. Or maybe the fact nothing is close to Claude for a long time made me too excited for this model haha
I asked them both to create a simple system. GLM excelled at design, presentation, and clean code. You can’t reach that conclusion by testing, because GLM is simply better.
Used GLM 4.5 air and 4.5 running locally, both were really solid. One thing - there might be some shenanigans with tool calling on these(locally with vLLM on release tool calls weren’t working till some later bugfix), some report that they format tool calls differently, so keep an eye on that.
Yes please!!! Replace Kimi K2 with GLM 4.5 if you guys want to.
I know Fireworks is a Cursor partner. GLM 4.5 is cheaper ($0.55/$2.19 vs $0.6/$2.5) and provides around 110-120 tok/s on Fireworks which is roughly double the speed of Sonnet 4 Thinking
Here we go again with the “New state-of-art free premium cheaper model better than Claude” of the time 
2 Likes
IMO: I think there’s more to it than just the model itself… I find Cursor to be the best AI-powered IDE, but I’ll admit that from time to time, I get frustrated because I don’t have access to new models as quickly as with some open-source alternatives (like Roo Code). I also wonder if most people would appreciate having reliable support for OpenRouter. Maybe I’m completely wrong, but the truth is that sometimes models take a long time to be integrated or maybe it’s just that we’re eager to test these so-called “revolutionary” models with the full power of Cursor.
It’s gettin close. Would def be a step up for this fckd auto mode.
Cursor has the opportunity to make AUTO mode work well. Customers will be happy and costs will fall. With the release of high-quality open-source models for coding, such as Qwen-3 Coder, Kim K2, GLM, Qwen-3, or even older models like GPT-4.1.
Cursor can now replace the current models in AUTO mode with these models. They are infinitely cheaper than Sonnet, users would be happy with the quality, and consequently would have less need to use Sonnet or other more expensive models. Everyone would win.
absolutely. Would be nice to be even possible to use alternative models after they dropped the price bomb, but Kimi-K2 is still not usable and Qwen3 is not even selectable.
Claude is still the best imo, but while Qwen3 and Kimi don’t really come close, this is actually usable. Definitely close enough to Claude quality to use for simpler prompts. Claude is so expensive especially now that Cursor is charging API prices pretty much 
. Even being able to pay it, it feels like a waste to spend so much on Claude when this gets 90% of the way there for much cheaper.