Anthropic has disclosed that it is already using an unreleased artificial intelligence system called Model 2, describing it as somewhat more capable than Claude Mythos 5. The previously undisclosed model was mentioned in the company’s August 2026 Risk Report, which covers developments up to July 15. According to the report, Model 2 was one of three frontier or near frontier AI systems that had not been publicly released. While Anthropic has confirmed that the model is actively being used within the company, it also stated that there are currently no plans to make it available to external users.
The strongest indication of Model 2’s capabilities comes from CoBench v2, Anthropic’s internal benchmark that evaluates AI performance using 449 real research and engineering problems previously solved by the company’s own employees. On this benchmark, Model 2 achieved a score of 62.8 percent, outperforming Claude Mythos 5, which scored 50.3 percent, and Mythos Preview, which reached 54.8 percent. Older models recorded significantly lower results, with Claude Opus 4.7 scoring 27.4 percent and Opus 4.6 reaching 15.6 percent. Despite this improvement, Anthropic estimates that an AI system would need to achieve at least 85 percent on CoBench before it could fully replace the work of its technical research staff. This means Model 2 remains well below that level, although the company considers it a noticeable improvement for many internal technical tasks. Anthropic also noted that the performance gain from Mythos 5 to Model 2 is smaller than the earlier improvement observed between Claude Opus 4.6 and Mythos Preview.
Beyond benchmark testing, Anthropic confirmed that Model 2 is already being used extensively across a variety of internal operations. Alongside Mythos 5, the model supports coding, data generation, research, engineering work and other agentic workloads. The company also uses the system for persistent AI agent deployments, highlighting its practical role in day to day technical activities rather than limiting it to experimental evaluation. Even with this level of internal adoption, Anthropic emphasized that Model 2 has not completed its full set of predeployment assessments. As a result, the company says it has lower confidence in its complete understanding of the model’s capabilities compared with AI systems that have undergone its standard evaluation process before release.
Anthropic also addressed questions about the model’s behavior and safety, stating that its internal deployment review did not identify any new or more concerning forms of misalignment beyond those already observed in Mythos 5. While the company continues to evaluate the system, it has not indicated that Model 2 will eventually be released as Claude Mythos 6 or under any other commercial name. At present, the company maintains that there are no plans for a public launch, leaving the unreleased model as an internal tool despite its stronger performance across selected technical tasks. The disclosure nevertheless provides a rare look into Anthropic’s ongoing AI development efforts and demonstrates that the company is continuing to build and test more capable models behind the scenes while following its internal evaluation and deployment process.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
