Sonar evaluated 53+ LLMs against 4,444+ Java programming assignments using SonarQube Enterprise to assess code quality beyond standard benchmarks like SWE-bench. Findings show that while models like Gemini 3.1 Pro High achieve 84% functional correctness, they produce significant security issues, bugs, and verbose code — GPT 5.4 generated over 1.2 million lines for the same assignments. Newer models tend to introduce subtler, harder-to-detect bugs. Sonar proposes an ACDC (Agent-Centric Development Cycle) framework with three stages: Guide (context augmentation and data treatment), Verify (SonarQube agentic analysis via MCP, running in 1-5 seconds before commit), and Solve (a remediation agent that auto-fixes issues and creates PRs). A public leaderboard at sonar.com/leaderboard tracks model quality metrics.