Anthropic has quietly been running its most capable AI model ever — and you’ll never get to use it. The company’s August 2026 Risk Report reveals an unreleased system called ‘Model 2,’ classified in the Mythos tier, that edges out Claude Mythos 5 on internal benchmarks by about 1.5 points on the AECI capability index. That’s a real gain, but nothing like the leap from Opus 4.6 to Mythos, which Anthropic itself seems to acknowledge by describing the improvement as modest rather than seismic.
The kicker is what Model 2 actually does inside Anthropic. The company leans on it heavily for coding, data generation, and research — sometimes through agents that run around the clock. Claude now writes most of the production code in Anthropic’s own systems, and Model 2 appears to be a big part of that pipeline. The model went through an internal review, though not as rigorous as the one Mythos 5 received, and Anthropic says it found no new or more concerning misalignments. Overall misalignment risk? Rated ‘low.’
So why keep it locked up? Anthropic says there are no plans to release Model 2 externally right now, and that’s the part worth pausing on. A frontier lab holding back its strongest model isn’t unprecedented — OpenAI has done similar things — but it does raise a question the industry keeps circling: if a model is good enough to run your own infrastructure, why isn’t it safe enough for everyone else? The report doesn’t answer that directly.
What’s clear is that Anthropic’s internal AI advantage is widening. The company isn’t just building models; it’s using them to build the next ones. That feedback loop — AI writing Anthropic’s code, generating its training data, running its research agents — could compound faster than anything we see from the outside. The public gets Mythos 5. Anthropic gets Model 2. For now, the gap between what a lab ships and what it actually uses is the real story.
