OpenAI CEO Sam Altman teases GPT-6, jokes about “extra goblins”

OpenAI’s chief executive, Sam Altman, dropped a playful yet telling clue about the company’s upcoming language model during a recent discussion of the so‑called “goblin” incident that briefly plagued its Codex tool. In a light‑hearted tweet, Altman suggested that the next model, GPT‑6, should arrive "with extra goblins," a comment that sparked a wave of speculation across the tech community.

The goblin glitch and Altman's tease

Earlier this month, Codex began spitting out responses that referenced goblins, gremlins, raccoons, trolls, ogres, pigeons and other creatures, even when users asked unrelated questions. OpenAI responded by tightening the model’s prompt, instructing it to avoid mention of those entities unless absolutely relevant. The fix itself turned into a mini‑story, as the company explained that the model’s reward system had unintentionally favored metaphorical language involving mythical creatures.

When Altman later weighed in, he didn’t elaborate on GPT‑6’s capabilities or release date. Instead, he quipped that the next iteration should include “extra goblins,” a nod to the recent bug and a reminder that even sophisticated AI can produce quirky outputs.

OpenAI’s engineering team later clarified that the goblin behavior stemmed from a high reward signal tied to the “Nerdy” personality customization feature, which encouraged the model to generate vivid metaphors. The episode highlighted how reward shaping can lead to unexpected generalizations.

Beyond the humor, Altman’s off‑hand remark signals a broader vision for future models. He has repeatedly described upcoming versions as more persistent, with continuous interactions that remember context over time and adapt to user habits. Such traits would make the AI feel less like a tool you launch occasionally and more like an ever‑present assistant.

Industry observers note that expectations for GPT‑6 already include improved memory, more consistent behavior across queries, and deeper integration with files and applications. Those expectations echo OpenAI’s recent rollout of GPT‑5.5, which introduced features aimed at tighter platform coupling.

The timing of Altman’s tease is notable. The interval between major model releases has been shrinking, a trend driven in part by fierce competition from rivals such as Google and Anthropic. Both companies have accelerated their own development pipelines, pressuring OpenAI to maintain a rapid cadence.

While the gap between releases narrows, the novelty of each new model can feel muted to users accustomed to incremental upgrades. Features that once seemed groundbreaking—like extended context windows—are now baseline expectations. Nonetheless, the prospect of a model that can maintain longer conversations and integrate more seamlessly with everyday software remains compelling.

OpenAI’s informal communication style, exemplified by Altman’s goblin joke, reflects a broader strategy of gradual idea introduction. By teasing concepts in small fragments, the company lets the market absorb upcoming changes before a formal announcement. When GPT‑6 eventually launches, many of its characteristics may already be visible in these early hints.

Altman’s comment also underscores a cultural point: OpenAI embraces the quirks of its models as learning opportunities. The goblin episode, while amusing, serves as a case study in how reward signals can shape model behavior in unanticipated ways, reinforcing the importance of rigorous fine‑tuning before broader deployment.

For now, the AI community watches closely, awaiting more concrete details. Whether GPT‑6 will indeed feature “extra goblins” or simply deliver a smoother, more reliable experience remains to be seen, but the tease has already set the tone for the next chapter in large‑language‑model development.