The capabilities of artificial intelligence are rapidly improving. From the early days when GPT-3.5 could only handle simple questions and casual conversations, to Claude 3.5 enabling quick coding, to DeepSeek meeting the minimum requirements for storytelling, and Doubao achieving human-like voice and fast emotional responses for virtual companions, and most recently, GLM 4.5 delivering impressive programming power.
Now, models like GPT 5-nano and GLM 4.5 have already surpassed Claude 3.5 in capability while maintaining affordable pricing, which is truly exciting.
I’m curious—at this point, it seems that only Anthropic has yet to release a low-cost model. Is Cursor’s continued price increase actually due to Anthropic?
But are people really that attached to Anthropic?
For example, when I used only Anthropic’s Claude for coding, it was simply because Cursor initially allowed Agent capabilities only for Claude models. As the restrictions were gradually lifted, I just stuck with what I was used to.
Recently, when GPT-5 was free, I also had a good experience using it.
It’s clear that in the foreseeable future, AI will only get better and cheaper. Many smart people are working hard to optimize AI.
For instance, the founders of DeepSeek recently proposed the Native Sparse Attention (NSA) mechanism, which, through algorithm and hardware co-optimization, directly increased long-text processing speed by 11 times. Even more impressively, performance not only didn’t drop but actually surpassed traditional full-attention models.
With the release of DeepSeek R2/V4 and GLM5, I believe the cost for Cursor to achieve Agent-level model capabilities can be further reduced.
I actually have an idea: recruit some users as seed users (like me), and by default let them use low-cost models such as GPT 5-nano, DeepSeek V3, and GLM 4.5. If users are dissatisfied with the generated results, they can click to upgrade to a higher-end model.
Although this may be a bit inconvenient for users, as long as Cursor offers extra free periods, users will likely accept it.
By continuously training with users’ manual choices, accurately directing their needs to super, advanced, or standard models, “Auto” can truly become “Auto.”