Anthropic’s Model 2 is reportedly improving internal coding, agentic workflows and data-generation tasks, but its limited safety evaluations do not show a clear overall advantage over Mythos 5.
Anthropic is testing a more capable artificial intelligence system known as Model 2 internally, but the company currently has no plans to release the model publicly, according to its August 2026 risk report.
The company says Model 2 has delivered improvements in several internal applications, including software development, agent-based workflows and data generation. However, its safety evaluations present a more nuanced picture of the system’s capabilities.
According to Anthropic, Model 2 performed at levels comparable to or below Mythos 5 in the chemical and biological evaluations that were conducted. Anthropic also noted that its assessments of Model 2 in these areas were more limited, making it difficult to draw broad conclusions about the two systems.
Model 2 to Remain an Internal System
Anthropic has confirmed that Model 2 is not currently intended for external release. The decision does not indicate that development has ended. Instead, the company continues to use the system internally alongside Mythos 5.
Keeping Model 2 within Anthropic gives the company an opportunity to evaluate and apply its capabilities in areas such as coding, agentic tasks and synthetic data generation without making the model directly accessible to customers or independent researchers.
As a result, users currently have no direct way to test Model 2 through Anthropic’s public AI products.
Why “Stronger” Is Not a Simple Comparison
Anthropic’s report also highlights the difficulty of ranking increasingly capable AI systems with a single measure.
Model 2 appears to provide meaningful gains on some internal tasks, but the available chemical and biological evaluations do not demonstrate a corresponding overall advantage over Mythos 5. On the tests that were conducted, its performance was generally comparable to or weaker than the older system.
Anthropic further acknowledged that some task-based evaluations may no longer adequately capture broader advances in AI capabilities. As systems become more sophisticated, older benchmarks can struggle to measure their full range of abilities.
Limited Safety Testing Leaves Questions Open
The findings should not be interpreted as proof that Model 2 is either safer or less capable overall than Mythos 5. Rather, they offer a limited assessment based on the evaluations Anthropic has carried out.
Because Model 2 has not been released publicly, independent researchers also have fewer opportunities to examine its capabilities and safety characteristics.
The situation underscores a broader challenge facing the AI industry: safety evaluations must evolve alongside increasingly capable models. A system can show substantial gains in one area while producing very different results on specialized safety assessments.
For now, Anthropic appears to be treating Model 2 primarily as an internal system. While the company reports improvements in selected areas, its available risk assessments do not establish a clear overall capability advantage over Mythos 5. With no public release currently planned, much of what Model 2 can ultimately do may remain known only within Anthropic.
About The Author
