Wait I think there are 2 levels of summary. Anthropic is definitely not showing its real thinking even with enterprise agreements. For example in Claude.ai the thinking traces are not real and are themselves summaries.