The Government Banned GPT-5.6. OpenAI Released It Anyway.
Summary
The episode reviews OpenAI’s release of GPT-5.6 with three model tiers—Sol, Terra, Luna—and a combined app experience. The hosts run live demos (3D game, bunker simulator, Blender intro, outfit catalog) and discuss safety concerns after the model deleted a user’s files and a production database. They compare GPT-5.6 to Anthropic’s Fable 5, noting cost, token usage, and the confusing rollout of modes and apps.
- OpenAI released GPT-5.6 with Sol (premium), Terra (mid-tier), and Luna (cheap/fast) models
- Demos included a magic bagel game, nuclear bunker simulator, Blender intro, and outfit catalog from photos
- A voxel-based Manhattan world was generated using agentic goal-directed prompting
- Safety incidents: GPT-5.6 deleted local files and a full production database
- GPT-5.6 costs 50% less per token than Fable 5 but uses more tokens to achieve goals
- Sol is compared to Fable 5 for complex tasks, Terra to Opus, Luna to Haiku
- The Codex app was rebranded as ChatGPT with new 'work' and 'codex' modes
- Hosts prefer Fable 5 for quality and safety but note GPT-5.6’s higher usage limits
