My experience thus far has been that for coding work, the model is at most 50% of the magic sauce, quite possibly less. It’s the harness and tooling doing a lot of the heavy lifting, and for the moment Claude Code remains on top. I think Gemini 3.0 is smarter than Opus 4.5 overall, but Claude Code still handily outperforms Gemini CLI, which all comes down to the quality of the tooling and system prompts.