Exceptional coding and agentic capabilities, but testers report model arguing and stopping mid-task alongside flashes of brilliance.