That's interesting, b/c that research from a few months ago seems to line up with the article. The article actually calls out 38x when you compare CC with Cursor's API token pool only. But then when you add in the Composer tokens, it becomes 5x. Obviously Composer's a different level of model than Sonnet/Opus.

I think that may be comparing cc against opus with cursor. This post compares against Auto+Composer pool. It does mention that if you used a frontier model then the gap would be even wider.