Anthropic is running an unreleased frontier AI model codenamed "Model 2" strictly behind closed doors, quietly outperforming every publicly deployed iteration of Claude. Disclosed in Anthropic's Risk Report from August 2026, the Mythos-class system is reserved entirely for internal R&D, synthetic data generation, and autonomous coding agents that now generate the vast majority of the code running inside the company's production stack.
Yet the benchmark deltas explain why this system remains vaulted. On Anthropic's internal capability index (AECI), Model 2 edges out Claude Mythos 5 by roughly 1.5 points—a modest delta compared to the jump from Mythos Preview to Mythos 5, and negligible next to the generational leap from Opus 4.6 to Mythos. Anthropic's own documentation concedes the architecture is merely "slightly stronger overall" while underperforming Mythos 5 across several specific evaluations, pointing directly to diminishing marginal returns on standard scaling.
Keeping this frontier asset in a private loop highlights a calculated strategic shift. While Model 2 cleared internal safety gating with an overall "low" misalignment score, Anthropic deliberately bypassed the expensive red-teaming and compliance gauntlet required for commercial release. In an environment where regulatory scrutiny is mounting and incremental public gains invite disproportionate liability, burning compute exclusively to accelerate internal engineering yields far better leverage than another margin-eroding public API endpoint.