[DigitalToday intern reporter Seungah Yoo] Anthropic said it is operating an internal artificial intelligence (AI) model, Model 2, that performs better than the top model released to the public. It said it does not plan to launch the model externally.

On Aug. 17 (local time), online media outlet Gigazine reported that Anthropic disclosed the results of safety reviews of Claude Mythos 5 and Model 2 in an AI risk report released on Aug. 15.

In the report, Anthropic described Model 2 as a model with "slightly higher capabilities" than Claude Mythos 5. It said the model is widely used internally alongside Claude Mythos 5. For public service, it said it has released a version with additional safety measures under the name Claude Fable5, and it specified it has no plan to make Model 2 public.

Performance assessments were similar in internal user perceptions. Anthropic employees evaluated Model 2 as "clearly improved over Claude Mythos 5 in many tasks." But the improvement was limited. Anthropic said it was not a big leap like the shift from Claude Opus 4.6 to Claude Mythos Preview, the base model for Mythos 5, and that overall it was only slightly more capable.

In the internal benchmark CoBench v2, Model 2 was narrowly ahead. But patterns in the safety evaluation details were not simple. In a test in which an AI carries out a secret mission without being detected by a monitor, the success rate of Claude Mythos Preview with extended reasoning turned on was presented as 6 to 10 percent. Model 2 was slightly higher than Claude Mythos 5, but it was far lower than Claude Mythos Preview.

The disclosure draws attention in that it shows both Anthropic's product structure and its internal development line. The highest-performing model currently known externally is not Claude Mythos 5, but Claude Fable5, which applies additional safety measures. That means even if a higher-performing internal model exists, Anthropic decides on disclosure based not on performance but also on the level of safety control.

The report also mentioned the name Model 1. Some users, accordingly, speculated that Model 2 is the successor to Model 1 and that smaller models, Claude Mythos 5 and Claude Fable5, may have been distilled based on those teacher models and Claude Mythos Preview. Anthropic did not confirm whether that estimate is correct.

In this situation, the report shows that Anthropic internally operates a separate line of high-performance models apart from the public models, while limiting external releases to models that have undergone reinforced safety measures. A future point to watch is whether Model 2 remains for internal use only or, after additional safety measures, leads to a public model under another name.

As part of our Responsible Scaling Policy, we publish regular Risk Reports. These share detailed information on the risks of our systems and how prepared we are to address them. Our second Risk Report is now available: https://t.co/NgWnDmXZD3