Anthropic published a Frontier Red Team assessment on September 10, 2026, titled “Measuring tactical intelligence targeting and conventional weapons capabilities of AI models.” The assessment described evaluations of intelligence-targeting tasks, including account linkage, individual classification and geolocation, along with simulated conventional-weapons tasks involving drone terminal guidance, payload delivery and navigation under GPS denial or spoofing.
Anthropic reported that Opus 5 struck its target in 80% of launches in the easiest scenario, which involved a parked target with high visibility. Across 540 launches in nine settings, the model struck the target in 20% of launches, according to the company.
The evaluation did not directly establish real-world battlefield performance. The drone-weapons tasks were conducted in simulated environments, the account-linkage tasks used synthetic social-media content, and the report identified limitations in how closely its targeting scenarios represented realistic operations. The findings therefore describe performance in the tested setups rather than demonstrated capability in an actual battlefield or deployed weapons system.