Anthropic Tests More Capable Model 2 Internally, But Has No Public Release Plans

Anthropic has no current plans to release its Model 2 AI system to the public, even though the company says it is using the model extensively for internal work.

The AI company describes Model 2 as an improvement for several internal tasks, including coding, agent-based work, and data generation. However, its latest risk report suggests that the picture is more complicated when it comes to specific safety-related capabilities.

According to Anthropic’s August 2026 risk report, Model 2 performed at levels comparable to or below Mythos 5 in the chemical and biological evaluations that were conducted. The company also noted that these evaluations for Model 2 were more limited.

This means that saying Model 2 is simply “stronger” than Mythos 5 would not provide the full picture. The improvement reported by Anthropic mainly relates to internal tasks, while the available safety evaluations do not show a clear overall capability advantage.

Model 2 Will Remain Internal for Now

Anthropic has confirmed that Model 2 is not currently planned for external release. The company has not said that development of the model has stopped. Instead, it continues to use the system internally.

Model 2 and Mythos 5 are reportedly being used for tasks such as software development, agentic workflows, and generating data. These internal applications give Anthropic an opportunity to use the model without making it available to customers.

The decision also means that users will not be able to directly access Model 2 or evaluate its capabilities through public products at this stage.

See Also: Anthropic Will Now Watermark Every Word Claude Writes

What Does “Stronger” Actually Mean?

One of the key points in Anthropic’s report is that AI capability cannot always be measured using a single benchmark.

Anthropic says Model 2 offers a noticeable improvement on certain internal tasks. However, its biological and chemical risk evaluations produced a different picture. On the evaluations that were performed, Model 2 was generally comparable to or weaker than Mythos 5.

The company also warned that some of its task-based evaluations are no longer keeping up with broader improvements in AI capabilities. As models become more capable, older tests may not be detailed enough to measure their full abilities.

This makes it difficult to use the available evaluations to create a simple ranking between Model 2 and Mythos 5.

Safety Assessment Remains Limited

Anthropic’s findings do not mean that Model 2 is safer than Mythos 5. They also do not prove that it is less capable overall.

Instead, the report provides a limited view of the model based on the evaluations Anthropic conducted. The company has not released Model 2 externally, so independent users and researchers also have limited opportunities to test it.

The report’s findings highlight a wider challenge for AI companies. As models improve rapidly, safety tests must also evolve to keep up with new capabilities.

For now, Anthropic appears to be treating Model 2 primarily as an internal system. While the company sees improvements in some areas, its available risk evaluations do not establish a clear overall advantage over Mythos 5. With no current external release planned, more information about Model 2’s capabilities may remain inside Anthropic for the foreseeable future.