Anthropic’s AI Models Unearth Smart Contract Vulnerabilities

Anthropic has unveiled a report exploring the performance of its models, Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5, on the SCONE-bench benchmark, which encompasses 405 real-world exploited smart contracts spanning from 2020 to 2025. These models discovered exploitable vulnerabilities in these contracts, potentially costing attackers up to $4.6 million. Further analysis revealed that Sonnet 4.5 and GPT-5 identified two new zero-day vulnerabilities during a simulated test on 2,849 newly deployed smart contracts with no known security flaws. Notably, the API costs for GPT-5 were $3,476 for this simulation.