SWE bench is weird because Claude has always underperformed on it relative to other models despite Claude Code blowing them away. The real test will be if Gemini CLI beats Claude Code, both using the agentic framework and tools they were trained on.