Anthropic disclosed in its August 2026 risk report the existence of two internal-only AI models: Model 1 (comparable to Claude Mythos 5) and Model 2 (notably more capable). Model 2 is used heavily for code generation, research, and security testing but will remain internal. The discovery prompted Anthropic to reassess model risk from "very low" to "low" and raised concerns that their evaluation methods are falling behind the pace of AI capability improvements.