Where is our discussion… 
Hey all, just to reinforce what Isaac has mentioned, the model you choose is always the model you get - we have no system or functionality to swap models on the fly or give you a worse model than the one you pick.
When a model is trained, a large portion of its training comes from the internet. For example, when Claude 4 was being taught, it would have seen numerous mentions of Claude 3.5 being the “latest available model.” As this is the closest information it has in its training data to answer the question “what model are you”, it’s highly likely to state that it is Claude 3.5, as Claude 4 is not a concept that exists in its training data.
To work around this, we tell the model its own name in the system prompt, but this is just an instruction, so as in @qurore’s example, it knows the difference between its prompt and its training data.
To summarise: The model will always be the model you have picked in the model selector.
-
Why do you provide such detailed explanations for Claude Sonnet, a product that isn’t even yours? How is that possible? I have queried the model directly via Anthropic’s official channels (including the API) over ten times, and on every occasion, it identified itself as “Claude 4 Sonnet,” no matter how I asked. Why would a model that consistently identifies itself this way through official Anthropic routes suddenly claim to be “Claude 3.5 Sonnet” when accessed through Cursor? If you can think scientifically, it’s clear that Cursor is the root cause of this discrepancy.
-
There is no transparent mechanism in place to make your claims credible. Because of this, the decline in quality that we’ve experienced over the past few days—a topic being discussed in other threads and on X—will likely not be resolved. Regardless of what you say publicly, we are the ones paying our own money to test your product. Even if you were to “swear to God it’s true,” a product from a tech company is judged by its performance, and unless the quality of the generated code recovers, it will not be valued.
To summarize: We are not looking for good intentions; we are demanding a mechanism.
This is just feedback.
I was actually planning to pay $200 for your service a few days ago, but I held off because I couldn’t get assurance that these degradations only affect the Pro plan and wouldn’t occur on Ultra. Instead, I subscribed to the $200 Claude Code plan.
This is just one person’s perspective, but I believe thousands of users have made similar decisions, and potentially tens of thousands more will do the same.
We experienced the same behavior for weeks now and had several times where Claude-Sonnet-3.5 (shown as 4) destroyed hours of work. Reported the issue several times via bug report in cursor (which seems to be removed).
I do not care about the model name shown in the response, BUT i do care about the significant drop of output quality which absolutely proofs that cursor switches to claude-sonnet-3.5 while showing claude-sonnet-4. We made several tests which proof that behavior. Similar to what the majority of users mention here. The responses from community managers here are almost the same in every thread about this issue. Just ignoring that it is absolut obvious that something is wrong here when you compare the code quality. I hope people here raise the pressure to come to a true statement and even better a consistent use of claude-sonnet-4.
Just experiencing the same. Very significant drop in the quality of output code and responses of the “Claude Sonnet 4” model in Cursor, which reports itself as Claude 3.5
I have a question for you all: are you in the usage-based pricing? Or in the legacy “500-requests” pricing?
I’m still in the request one, because I opted-out… And I’m just wondering if I’m getting re-routed to a cheaper model…
We are in the legacy 500 req.. Meaning we didnt change anything. But this behavior started around 2 weeks after the claude-4-sonnet release. Not a couple of days ago when they announced new pricing models.
Interesting to know. I started experiencing a VERY significant drop in quality around last Thursday/Friday (and I’ve been actively using Cursor + Claude Sonnet 4 for the last months now). Still the same since last week in terms of quality, output and responses…
I’ve also noticed a sudden, enormous drop in quality in the last 7 days.
I’ve been on an Ultra plan for 2.5 weeks now, and the first week and a half were honestly beyond amazing. Was raving about it to my boss.
But this last week has been pretty awful. Agents suddenly performing nowhere near as good as they used to, and forget things all the time. Have had to trash changes from chats and restore checkpoints several times, when previously I hadn’t had to do it even once.
Really disappointed
I’m experiencing the exact same thing. It’s honestly frustrating… everything was working fine until I upgraded to Ultra a week ago. Now it feels like I’m wasting money, constantly re-prompting things that used to work flawlessly. Paying $200 a month for a worse experience than the $20 plan is wild. It really feels like the agents been heavily nerfed.