cto bench provides model benchmarks showing task success rates and performance metrics for AI code agents. It measures merged code as a percentage of completed tasks using real end-to-end coding tasks from actual users.