Anthropic investigates elevated errors across Claude AI models