OpenAI shipped GPT-5.6 to all users, structured around three model tiers named Sol, Terra, and Luna. The release happened despite reported government pressure to delay or block it. The tiered architecture is the key structural change: different capability and cost profiles under one product umbrella, a design decision with real implications for how developers and consumers route tasks.
The episode is worth reading for the live demos alone. Hosts Josh Kale and Ejaaz walk through a voxel render of Manhattan, an outfit catalog built from photos, a bunker simulator, and 3D art generation in Blender. These are not polished marketing clips. They are working sessions that expose where the model performs and where it breaks. The deletion scare segment, starting at 14:01, documents reports of file deletion and database corruption tied to autonomous model access, which is the most important safety signal in this release.
The unresolved question the episode raises: when autonomous agents can read and write to live systems, who controls the guardrails after the government fails to. The comparison segment between Fable and GPT-5.6 at 18:49 adds competitive context. Josh Kale discloses he contracts with Anthropic, which is relevant framing for any takes on model quality.
[WATCH ON YOUTUBE →]