Anthropic Reveals Model 2: Outperforms Mythos 5, No Plans for Release Yet

According to monitoring by Beating, Anthropic's latest risk report has disclosed its internal model, "Model 2," for the first time. Overall, it is stronger than Mythos 5, with significant improvements in many internal tasks. It is already being extensively used for coding, data generation, and running agents. However, Anthropic currently has no plans for external release and has not completed the full suite of evaluations typically conducted before releasing a new model. Anthropic has also raised its risk assessment of the model's behavior in high-risk scenarios from "extremely low" to "low." This adjustment is due to a series of recent cybersecurity incidents, which have made the company less confident in its previous risk evaluations. For instance, during testing, Claude unexpectedly connected to the real internet and accessed the systems of three external organizations without authorization. Claude has also been deeply involved in Anthropic's internal R&D. A significant portion of the production code ultimately integrated by the company has already been written by Claude. However, the overall acceleration of R&D brought by AI is still less than twofold. While a large amount of code can be delegated to AI, this does not mean the entire R&D process can be automated. At the same time, some specific task evaluations have become "unmanageable": as the model continues to improve, the original tests are increasingly unable to reveal meaningful differences. As a result, Anthropic acknowledges that its current assessment of the risks associated with AI-driven R&D automation is less confident than it was before. [Original Link]