OpenAI CEO Sam Altman Hints at GPT-6 Development Following ‘Goblin’ Bug Fix

Quick Facts

• Sam Altman posted about starting GPT-6 training with ‘extra goblins’ after OpenAI fixed a Codex bug
• OpenAI completed pre-training for GPT-5.5 on March 24, 2026, and released it April 23, 2026
• Enterprise revenue now represents over 40% of OpenAI’s business and is on track to reach parity with consumer revenue by end of 2026

OpenAI CEO Sam Altman dropped a playful hint about GPT-6 development after the company fixed an unexpected bug in its Codex tool that generated goblin-themed responses.

The issue arose when OpenAI’s Codex began producing unexpected creature-themed content. The company traced the bug to training rewards given during personality customization features, particularly the ‘Nerdy’ personality setting. OpenAI engineers discovered they had unknowingly assigned high rewards for metaphors involving creatures, causing goblin references to spread throughout the model.

OpenAI fixed the issue by adding specific prompts telling the AI to avoid discussing ‘goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.’

Altman joined the conversation by posting a screenshot with the caption: ‘Start training GPT-6, you can have the whole cluster. Extra goblins.’ While the comment was humorous, it marked the first public reference to GPT-6 development from OpenAI leadership.

The company recently completed pre-training for its model codenamed ‘Spud’ on March 24, 2026, at the Stargate data center in Abilene, Texas. This model shipped as GPT-5.5 on April 23, 2026, keeping OpenAI within the GPT-5 family.

AI researchers expect GPT-6’s release in Q2 2026, with May or June as the most likely months. OpenAI’s enterprise focus has intensified, with Fidji Simo, CEO of Applications, telling staff the company is ‘orienting aggressively’ toward high-productivity use cases.

Enterprise revenue now exceeds 40% of OpenAI’s total revenue and is projected to reach parity with consumer revenue by the end of 2026. The shift comes as 78% of organizations plan to increase AI spending, according to Futurum Group’s AI Platforms Decision Maker Survey of 820 participants.

GPT-6 is expected to feature long-term memory as its headline capability. Altman has described current AI memory as being in the ‘GPT-2 era’ relative to what’s coming, suggesting significant improvements in recall of preferences, projects, and conversations across weeks or months.

The next-generation model will also expand agentic capabilities with better goal decomposition, more tool integrations, and greater autonomy. OpenAI has shifted to reinforcement learning as the core training approach for GPT-5 and GPT-6 generations, moving away from pre-training paradigms used in earlier models.

This training methodology allows models to discover patterns and solve problems through trial and error rather than just absorbing existing text. OpenAI describes this as enabling AI to potentially contribute to new algorithms, physics, and biology discoveries.

Read more: Sam Altman just dropped a big hint that GPT-6 is coming soon