Fully agree. GLM-5.3 Flash is the first model that has made me seriously consider moving off since I can likely get a much better deal through Codex + GLM-5.3 Flash. Composer 2.5 was great until GLM-5.3 Flash exposed it.
I’m sticking around for a while longer because the Grok releases have impressed me and Grok 4.7 is coming up, but if we can’t get something to compete in the high-volume areas that GLM-5.3 Flash covers (especially with the OpenAI departure in November and upcoming Astra release), then I’ll have a pretty serious decision to make at that time.
As it stands, I’ve already been fairly frustrated with the known issue of not being able to spin up subagents of a family when using an API key for that family (except when the parent agent is the same family). That issue also prevents me from just using GLM-5.3 Flash through an API key, as I’d want to spin it up as swarms while still using Grok as my parent agent.
You’re definitely on to something. GLM-5.2 was added about a week since its release, and Kimi K3 within a few days. That’s proof they’re 100% ignoring GLM 5.3, GLM-5.3 Flash, and Deepseek v4 Flash 0731. It’s clear that the criteria is “don’t add open-weight models which have better price:performance than the 1st-party models”. GLM-5.3 Flash and Deepseek v4 Flash 0731 directly hit the value proposition of Composer 2.5, while GLM-5.3 hits the value proposition of Grok.
My prediction: We’ll never get GLM-5.3 Flash and instead we’ll get a Composer 3 (or Grok equivalent if they start doing tier splits like OpenAI and Anthropic), and we’ll probably also never see Deepseek v4 Flash 0731 - though that one has a chance since it’s worse than GLM-5.3 Flash. I also predict once Grok 4.7 drops, then we’ll suddenly see GLM-5.3.
The worst part? I’m actually fine with all of that as long as it happens before November. The Astra release was relatively underwhelming (in very specific aspects - though it’s obviously now the best in others), so Grok 4.7 has a good shot if Musk’s statements are accurate. That means all we’d need is something to compete in the high-volume low-end where GLM-5.3 Flash is.
All that said, I want to make clear: the actual proper solution is still for them to just add these models we’re requesting - which we can clearly see they’re capable of doing quickly over the course of their history, and even as recent as GLM-5.2 and Kimi K3.
Just to add some context here: evaluating and onboarding a new model is a real process on our end, not a switch we flip. GLM 5.3 is definitely on our radar, but we don’t have firm plans to share just yet.
Historically, we’ve added support for open-source models such as K2.7/K3 and GLM 5.2. It doesn’t always happen immediately, and it gets prioritized alongside everything else we’re working on! There’s no deeper motive at play beyond that.
Voicing support here is helpful as it gives us a clear signal of demand and helps move things along, so please keep it coming.
Thanks for the response, and we definitely appreciate the transparency, at least now we’re in the loop.
While i made this post as a request for GLM 5.3 i truly am looking for general latest models, especially those that rank top on Intelligence/cost ratio like GLM 5.3 Flash and now with the Muse Spark 1.3 as well, that said, I’ll be honest that the longer this gap stays open, the more it nudges people like me toward just running these models through other tools where they’re already available and I’d much rather stick around and keep building on Cursor than switching, so hopefully this thread (and the demand behind it) helps make the case internally that it’s worth prioritizing sooner rather than later.
I accidentally got a 5-day weekend over Labor Day, so I’m spending this entire time familiarizing myself with T3 Code + OpenCode + OpenRouter so that I can test out Muse Spark 1.3, GLM-5.3 Flash, and Deepseek v4 Flash 0731. While they’re certainly not as impressive as frontier models, this specific model combo + Gemini 3.8 Flash covers probably 80-90% of my use cases while still being cheaper than Cursor (though I’ll admit that’s primarily due to Muse’s Contributor variant, which is fine for my use cases).
Cursor’s 1st-party subsidies are incredibly generous, but not having a high-volume competitor in the low end is offsetting that to a pretty significant extent. Composer 2.5 was great until it wasn’t.
Well… It was all good before Cursor was bought… It seems now we take a peek on the future “add new models” decision process. Nevertheless we can always change providers (I’m from one of the many dev teams that leave Claude or Copilot) . if they get into that nasty behavior, we know the answer.. So all is fine…
Granted it wouldn’t use Cursor’s included usage, but you could easily set this up as its own model. You would pay for it separately, via like OpenRouter. But it might still be worth the cost/benefit tradeoff.
