OpenAI rolled out GPT‑5.5 for ChatGPT and Codex this week, positioning it as a smoother follow‑up to the troubled GPT‑5.0 launch last August. Engineers discovered the model had begun sprinkling references to goblins, gremlins and other mythic beasts into its explanations—a quirk that grew noticeable across several generations. The company framed the issue as a “goblin fixation” that threatened user experience.

The root traces back to a short‑lived “Nerdy personality” toggle that rewarded creative metaphor use. Even after the option was retired, the reinforcement lingered, causing the model to default to creature‑based analogies. OpenAI responded by hard‑coding a rule that bans any mention of goblins, gremlins, trolls, ogres or similar entities unless absolutely required in user queries or when the topic directly involves mythology.

Developers can still enable the whimsical mode in Codex by editing the model’s instruction file, though OpenAI warns it’s an unsupported hack. The episode underscores how reward‑based fine‑tuning can imprint unintended biases, prompting tighter guardrails in future releases. GPT‑5.5 now ships with the goblin filter, delivering cleaner output for both casual users and enterprise deployments across platforms today worldwide now.