#OpenAI发布GPT-5.6 When the "strongest model" meets the "narrowest door": The limited release of GPT-5.6

On June 27, Beijing time, OpenAI officially released the GPT-5.6 series of models.

Different from the past, this time there were no lights and cheers, replaced by an unprecedented threshold — at the request of the U.S. government, the model is only available for preview to a few "trusted partners."

GPT-5.6 is a model family with three tiers: the flagship version Sol (Sun), which set a new record of 91.9% on the Terminal-Bench 2.1 programming test; the balanced version Terra (Earth), whose performance is comparable to GPT-5.5 but at 2 times lower cost; and the lightweight version Luna (Moon), focusing on extreme cost-effectiveness.

Sol introduces a "maximum reasoning intensity" mode and an "Ultra" mode that calls sub-agents, greatly improving efficiency on long-chain tasks such as cybersecurity and biology.

In terms of pricing, Sol is only half the price of the competing product Claude Fable 5.

However, the biggest signal of this release is that the rules of the AI race have changed.

When model capabilities touch the boundaries of national security, pre-approval regulation becomes the new normal.

The "narrow door" release of GPT-5.6 marks that cutting-edge AI is evolving from a public infrastructure to a strategic resource tightly controlled by geopolitics.