GPT-5.4 tops short-story benchmark

This is a useful signal for GPT-5.4’s creative-writing quality, but it is still a niche benchmark result rather than a definitive “best model” verdict.

• –The ranking change comes alongside a new pairwise evaluation method, so the methodology shift is part of the story, not just the model jump
• –Creative-writing wins can matter for developers building agents, roleplay, marketing, or long-form generation workflows where voice and coherence matter
• –Because the result surfaced via Reddit and X rather than a formal benchmark paper or leaderboard site, it should be treated as directional community evidence
• –OpenAI’s same-day GPT-5.4 release gives the result extra weight, but cross-run comparisons deserve caution until the new grading mode settles