The only LLM to beat Gemini 3.1 Pro as of today in a coding-related benchmark is Codex, specifically, GPT-5.3-Codex, with the Extra High reasoning effort enabled.