Codex Candy Eval Benchmark

Skill by ara.so — Codex Skills collection.

A Python-based benchmarking tool that evaluates OpenAI Codex/GPT models using a standardized candy math puzzle. Measures reasoning token usage against answer correctness to assess model performance degradation patterns (降智测试).

What It Does

• Batch tests Codex CLI with a specific candy distribution math problem
• Tracks reasoning token consumption across different effort levels
• Calculates accuracy rates by detecting the correct answer (21)
• Supports multiple reasoning effort levels: low, medium, high, xhigh
• Zero third-party dependencies - pure Python with Codex CLI integration

Prerequisites

You must have Codex CLI installed and authenticated: