Does anyone know what the base model for composer 2 is? Seems like GPT-4 ?
Also, am I the only one who hits 100% api usage and 0% auto+composer usage every month?
Bit of back story:
I ran out of API tokens and decided to play around with composer 2 on a new project to see what the quality is. It was great at first during planning mode and initial development (Roughly ~300 iterations / AI runs) I didn’t start running into issues until the code base exceeded ~100K lines. Where I’m at now it doesn’t seem like it can make any sense of the code base that it designed and made.
For example, I just made a plan to go through do some debugging work, correct some issues, then deploy & test code on a local test instance. At the end of the planned work it had no clue what it even just did. Then I started talking to it more and it started telling me to do things manually that were defined as automatic with supported documentation, flows, etc from the start of the project. I haven’t seen this sort of behavior since early GPT5 and late GPT4 models where it was almost a necessity to keep a code base < 100K lines.
I do have to commend Cursor on the efforts put in to training for the planning mode. The composer 2 is better than any others I’ve tried at generating planned actions.