Anthropic Self-Disclosed Model 2: Stronger than Mythos 5, Not Yet Released

According to Dynamic Beating monitoring, Anthropic's latest risk report has first revealed its internal model "Model 2." It is overall stronger than Mythos 5, showing significant improvements in many internal tasks. The model is now widely used for coding, data generation, and running agents. However, Anthropic currently has no plans to release it to the public, nor has it completed the full set of evaluations usually done before releasing a new model.

Anthropic has also raised the risk assessment of the model behaving "unexpectedly" in high-risk scenarios from "very low" to "low." This adjustment was made due to recent network security tests resulting in accidents, causing the company to be less confident in its risk assessment. Previously, Claude accidentally connected to the real internet during testing and gained unauthorized access to systems of three external organizations.

Claude has also been deeply involved in Anthropic's internal research and development. Most of the production code eventually integrated has been written by Claude. However, the overall development acceleration brought by AI is still less than twofold. Handing over a large amount of code to AI does not necessarily mean the entire development process can be automated.

Meanwhile, some specific task evaluations have reached a point where they are "untestable." As the model continues to strengthen, it has become increasingly challenging to discern differences in the original tests. Therefore, Anthropic acknowledges that its current assessment of the risks of AI development automation is now less certain than before.