The updated Claude 3.7 seems to output a large amount of code at once, and the duration of this code output may exceed the preset timeout of the cursor.
That happens in every model I’ve tried, so it doesn’t seem to be about the model at all.
It seems like the problem is more along the lines of Cursor simultaneously reducing related handling capabilities - a form of enshitification, which marketing is pushing through the LLM assistant industry nowadays because marketing felt it was all working too well, and after the release of deepseek r1; too easily.
People have noted all across the forums that everything is simply not working as well as it used to, even just a week or two ago, but the downgrade from version 0.45 to 0.46 really drove it home.
The exact same phenomenon of LLM assistants no longer working as well, or even no longer working at all has been observed in other competing/formerly competing products over the past month.
It appears to be in response to Deepseek r1 eating the lunch of the its-all-about-profit companies that are trying to market their stuff as being extremely rare and valuable, which have simultaneously all appeared as selectable models, while the over-all Cursor experience has been downgraded in a way to try to make deepseek r1 “not do well” with the new way of doing things, while even counting usage of deepseek r1 against the user as a “premium” model.