inside Claude Code's leaked source code, there are some very INTERESTING bits about their next model Capybara, next features, and Claude's internal tricks: > with UNDERCOVER mode the claude team could secretly contribute to public repos > Anthropic flags when you're rude to Claude (probably for training) > in-app voice mode, and ultra-plan (up to 30 minutes planning are next features) and so many unreleased but implemented features exist too. > if Claude realizes you're trying to distill it, it gives you fake responses > for some sub-agents and processes, Claude silently switches from whatever model you're on to Haiku for speed. > their next model Capybara is explicitly prompted to avoid over-commenting (basically shut up unless necessary) > Capybara v8 has a false claim rate of 29-30%, double the FC rate for a previous v4 version. probably an internal benchmark. it's explicitly prompted to "Report outcomes faithfully".