The chatgpt-4o-latest endpoint has branched away from the 4o-08 (and/or 4o-11) models quite a while ago, so it's not unlikely they've updated it again.

Sonnet has also had silent changes, likely at the system prompt level - not the public one, the non-public "alignment" one that is even applied to the API - that have made it lazier in terms of maximum output length (despite max_tokens values).