The Model War's Decider Shifts from "Who's Strongest" to "Who's Allowed to Sell to Whom" — OpenAI shipped GPT-5.6 (Sol/Terra/Luna; 1.5M-token context, ~43% over GPT-5.5). Flagship Sol is OpenAI's strongest cybersecurity model — competitive with Anthropic's Mythos Preview on ExploitBench using only ~1/3 the output tokens; Terra gives GPT-5.5-level performance at ~half the cost; Luna is the cheapest. But at the US government's request, GPT-5.6 ships first to only ~20 government-approved partners. Judgment: this continues ENTRY 76 (Fable 5 pulled offline) and ENTRY 85 (passport wall), but evolves a step — Anthropic was shut down reactively (ship, then get pulled); OpenAI now does a "licensed release" proactively (get the government's approved-buyer list first, then ship). Release rights are normalized into the national-security frame — from incident to factory process. The model war's decider shifts from "higher benchmark / cheaper" to, at the top tier, "who can legally sell the strongest capability to whom." GPT-5.6 is born into two worlds — ~20 licensed partners run full Sol, everyone else a restricted version — the first "military-grade vs civilian-grade" legal access tiering, a layer deeper than ENTRY 83's telco-ification. Price war still rages at the civilian layer (Terra half-price, DeepSeek halving); the flagship contest moved from price to access. Primary-market read: (1) "model switchability" goes from best practice to survival need again (ENTRY 85) — single frontier vendor = inherit its nationality/list risk; (2) validates "regulation as moat" (ENTRY 84) — the moat is now "qualified to get the government sell-license," harder to copy than parameters; (3) Chinese open source gains again (ENTRY 76) — downloadable, on-prem weights gain a certainty premium when the strongest closed model ships by list/nationality. Rollback: if GPT-5.6 opens fully within weeks, downgrade "licensed release normalized" to "launch-phase limit."