They have a giant pile of data from every customer that didn't disable data retention

I keep hearing this but they’ve been bleeding users since Claude Code came out so a bulk of their data is pre-Sonnet 4 and I’m not sure how the data from users prompting weaker models will help them now?

They did use that data to make Composer 2.5 which was decent but still a step back from GPT 5.5 and Opus 4.8. Though it’s really good at UI.

	▲	connicpu 2 hours ago \| parent [-]
		Not sure how much data they retained but it's possible they also have the _input_ tokens, and in the "early" days (how far everything moves in just a couple years...) more of those input tokens would have been non-AI-generated