Claude Code, OpenCode face off on roguelite

The interesting part isn’t that both stacks got to a working game; it’s that they did it with different tradeoffs in time, tokens, and implementation style. This is a solid demo of agentic coding maturity, but it’s not a rigorous benchmark.

• –Same prompt, no follow-up prompts, and no manual fixes makes the harness and setup part of the story, not just the model
• –Opus 4.7 used more tokens and took longer in the post’s comparison, while OpenCode plus a local Qwen stack stayed lighter
• –The visual differences suggest “it runs” is too low a bar; spec adherence and iteration efficiency matter just as much
• –OpenCode’s model-agnostic terminal workflow keeps local/open-weight setups competitive against proprietary coding agents
• –This is useful signal for AI coding workflows, but still anecdotal rather than reproducible eval data