Anthropic is running an unpublished AI model internally that is more capable overall than its public Claude models, according to The Decoder, which cites the company’s August 2026 Risk Report. The model is referred to as “Model 2” and is not currently planned for external release, according to the report as summarized by The Decoder. The reported model sits in Anthropic’s Mythos class. The Decoder says the report describes Model 2 as slightly stronger overall than Claude Mythos 5, while still weaker in some areas. On Anthropic’s internal capability index, AECI, Model 2 is reported to score about 1.5 points above Mythos 5. That gain appears incremental rather than a major step change, based on The Decoder’s account of the report. The article says Model 2 does not represent a capability jump comparable to the move from Opus 4.6 to Mythos, and that the increase is smaller than the jump from Mythos Preview to Mythos 5. Inside Anthropic, the model is reportedly used heavily for coding, data generation, research and engineering. The Decoder says some of that use runs through agents that operate continuously. The same report also says Claude now writes most of the code in Anthropic’s production systems. The safety picture, as reported, is measured. The Decoder says Model 2 underwent an internal review before deployment, but was not tested as thoroughly as Mythos 5. Anthropic reportedly found no new or more concerning misalignment issues and rated the overall misalignment risk as “low.” For readers, the main constraint is sourcing: the cluster contains one reputable article, not multiple independent reports or a primary-source item. The details are specific and attributed to an Anthropic risk report, but they remain single-source within the provided material. Who benefits: Anthropic’s internal teams appear to benefit most, particularly in coding, research and engineering workflows. Public Claude users do not appear to have access to Model 2 under the current reported release posture. Who's exposed: The main exposure is informational: outsiders evaluating public Claude models may not see the full capability level Anthropic is using internally. The report also notes that Model 2 was not tested as thoroughly as Mythos 5, though Anthropic reportedly rated misalignment risk as low.