Is anyone else experiencing major quality issues with Claude models lately?

[JohannesBin (Johannes)]

Hey everyone,

I’m hoping I’m not the only one dealing with this, but I’ve been having some really frustrating issues with Cursor over the past few weeks and I’m wondering if others are seeing the same thing.

Background: I’m a heavy user, paying hundreds per month, and I’ve generally been happy with the service until recently.

What I’m experiencing:

1. Claude model quality has tanked - When I select Claude-4-Sonnet or Claude-4-Sonnet-Thinking, the responses feel like they’re coming from a completely different, much worse model. The quality is nowhere near what I was getting before. It’s like night and day.

2. The “thinking” model does not think - This is driving me crazy. Yesterday I used Claude-4-Sonnet-Thinking for an entire day of coding, and it only actually went into thinking mode ONE time. The rest of the time it just immediately started generating like a regular model. What’s the point?

3. I can see code being generated in real-time - This started happening recently and I’m pretty sure this isn’t normal? Like I can watch it type out code character by character, which seems like it might be a bug.

4. Bills keep going up - My monthly costs have been steadily increasing even though the service quality has gotten worse. I know there’s been discussion about pricing changes in the forum but i haven’t been following it closely.

Has anyone else noticed this? I’m starting to wonder if there are some serious backend issues going on, or if maybe I’m being routed to some cheap fallback model when I request Claude? It genuinely feels like I’m not getting the actual Claude models I’m selecting and paying for.

I really want to love this tool - when it works properly, it’s amazing. But lately it feels like I’m paying premium prices for budget service. I’ve already reached out to support but curious if this is a widespread issue or if it’s just me.

Would love to hear if others are having similar experiences or if anyone has found workarounds.

Thanks!

[JohannesBin (Johannes)]

UPDATE: Well, this is interesting. I just tested something and it confirms my suspicions. When I asked Claude directly what model I’m talking to, it responded that it’s “Claude 3.5 Sonnet, not the thinking model” - even though my Cursor interface clearly shows “claude-4-sonnet-thinking” selected.

So either:

1. Cursor is mislabeling models in their UI
2. I’m actually being routed to Claude 3.5 instead of Claude 4 when I select Claude 4
3. There’s some serious confusion in their backend about which models are being served

This is exactly what I suspected - I’m not getting the Claude 4 models I’m paying premium prices for.

Claude 4 answers differently:

[JohannesBin (Johannes)]

UPDATE 2: Only gets worse. I asked who the current US president is, and it said Joe Biden, mentioning it has a knowledge cutoff date. But real Claude Sonnet 4 knows Trump is the current president as of January 2025. So not only am I getting Claude 3.5 instead of Claude 4, I’m getting an OLD version of Claude 3.5 with outdated information.

[krazexs (Marco G)]

yes absolutely, also the coding lost quality, in June i created a fairly complex app and cursor was doing well, not it became a nightmare!! I got a better code experience with Sonnet 3.7 than 4.0!!

[ChuckJonas (Charlie Jonas)]

I noticed this two, but after testing in AWS bedrock running sonnet4, it also says that it’s 3.5