Claude Code beats local Qwen, but the gap looks smaller
Claude Code with Opus 4.8 was compared with running locally through Ollama. Both received the same joke-style coding request: build GTA6 and make no mistakes. The goal was not a serious game project, but a playful test of what can do.
The codehamr agent used for the local test was intentionally simple. It had no plugins and no MCP, so all available context went into the task itself. Opus 4.8 won, as expected, but the local 27b model felt closer than expected.
Both systems produced something that could loosely be called a city with cars, and both failed in obvious ways. Better prompting may narrow the gap further, but that takes time. The codehamr is free and .
Key points
- Claude Code with Opus 4.8 was tested against local through Ollama.
- Both tools got the same unrealistic coding prompt about building GTA6 without mistakes.
- The local codehamr agent ran without plugins or MCP.
- Opus 4.8 performed better, but the local 27b model seemed closer than expected.
- The codehamr project is , so others can inspect or try it.