“Anthropic aims to improve its AI models to reach human-level reliability, estimated at 80-90%, on certain tasks.”