Benchmark of 14 AI coding agents: success ranges from 100% to 0%

The author ran 14 agentic AI coding tools on the same task set with the same gemini-3.5-flash-lite model on a GPU-less PC. Cline passed 7/7, Aider, OpenHands, Kilo Code and OpenCode 6/6, while SWE-agent, Void and Roo Code scored 0%. On local qwen3.5:4b the leaders hit 67% and most tools hovered near 25%.

• Cline scored 7/7 on gemini-3.5-flash-lite, Aider and OpenHands 6/6
• SWE-agent, Void and Roo Code failed every attempt on the cloud model
• On local qwen3.5:4b Cline, Teaspoon IDE and Void reached 67%
• Roo Code's repo is archived and attempts got no response