Claude Opus and GPT-6-Sol Lead AI Code Security Rankings

Veracode, a company that tests code security, checked 18 AI models on code generation across four programming languages and four types of vulnerabilities. This is the biggest jump since July 2025. The average share of secure code rose from 56% to 62%. The share of tasks with a vulnerability fell from 44% to 38%. The leaders are Claude Opus 5.5, Anthropic's current flagship model, and GPT-6-Sol, a model from OpenAI. Both scored 73%, beating the previous record of 68% set by GPT-5.5 in July. Java remains the weakest language, with only 31% secure code, while Python reached 64%.

Veracode's security director made a sharper point: a mid-tier model at 60% plus a security check is more reliable than a top model at 73% without one. Even the best models still let a vulnerability through in almost four tasks out of ten. This means code review and scanning remain a required step for anyone shipping AI-written code, not a formality you can skip.