Are you evaluating gpt5-thinking on high mode, via API or Codex CLI on Pro tier? Just wondering what specifically you compared since those factors affect its performance and context