I don’t know if you guys changed it, but its capabilities have gone down a lot since it was first released, and it’s very underwhelming 
3 Likes
I concur, I’ve used it since Friday with many intermittent “connection” problems driving me nuts. It seemed on par with 3.7 most of Friday, but frankly now it has gone down hill. I miss 3.7 - it went rogue out of scope often but with sufficient shepherding was at least predictable and more “grad-level” than sonnet-4.
Check that Sonnet 4 Thinking is being used. I noticed this morning Cursor had renamed Sonnet 4 Thinking to Sonnet 4, and added a new Sonnet 4 Thinking model additionally since last week. You might need to go to Settings > Models and enable it.
Are you suggesting that Sonnet 4 Thinking will give better results. I too feel 4.0 is reverting to a lower IQ. It’s cycling through it’s mistakes and not really being creative in circumventing deadends.
I mean, when Claude 4 was first released by Cursor, it worked pretty well. I assumed it was the Think version. But in the last couple of days, Claude has become really dumb. I wonder if Cursor is adjusting something that’s causing Claude 4’s abilities to decline.
Could be…they may be forcing you to shell out $ to get some decent results. I can’t judge.
yes. degradation in model performance has almost exclusively been anecdotally observed to destroy outputs given a circus driven system prompt no one sees. have fun doing the E calculation on a stochastic nightmare you don’t even get control over. The more BS prompting they add to protect themselves in an ad hoc fashion, the more the system and motivations are betrayed to play hide and seek with nothing worth value.