“Anthropic believes that to fully understand AI model safety, models must be released as products for real-world use, as laboratory settings are insufficient.”