Anthropic’s Strongest Model is Kept In House
In the August 2026 internal risk report, Anthropic mentions that they have access to a model that’s better than their public models: Model 2. This model is already used to write most of Anthropic’s production code, even though it’s never been released externally.
In Anthropic’s August 2026 internal risk assessment report, they mention having access to a model called Model 2 which is stronger than all their public models. This model is currently being used within their production systems and for writing most of their code. It is approximately 1.5 points stronger than Claude Mythos 5 on their internal capability index (AECI).
Before using this model, they ran some safety checks and deployed it with less testing compared to Mythos 5. There were no novel cases of misalignment found, so they believe the risk to be quite low. They don’t intend to share this model publicly. So we see that even at the world’s top AI safety team, their strongest models are different from what they make available to others.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →
Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles