the Most Capable Intelligence No Longer Has a Public Version — Facts: on July 21 Google shipped three new Gemini Flash models, the notable one being Gemini 3.5 Flash Cyber (vulnerability finding/patching via CodeMender), available exclusively to governments and trusted partners in a limited pilot — while flagship Gemini 3.5 Pro still hasn't shipped after two months. Place it in a four-month arc: April 7, Anthropic released Claude Mythos (83.1% on the CyberGym vulnerability-reproduction benchmark vs Opus 4.6's 66.6%), for approved organizations only, with no public API, no pricing, no GA; OpenAI's GPT-5.6 entered a government-gated preview on June 26 (~20 government-vetted partners) and only went GA on July 9 after clearance. The most capable tier of AI is systematically withdrawing from public release. Judgment: (1) the structure is what matters — Anthropic's Fable 5 and Mythos are the same base model, differing only in system prompt and access: Fable is the public version with guardrails (cyber/bio queries auto-route to the weaker Opus 4.8), Mythos is the approved-orgs version with full capability and no such constraints, placed in a new tier above Opus; one engine, two versions, separated by an export license — the most capable intelligence no longer has a public version, and what the public can buy is always the valve-throttled one, fulfilling her June 13 call that Fable 5 would be treated as a munition / export-control target (ENTRY 76); (2) the hardest evidence: on June 12 the US Commerce Secretary sent Anthropic an "Is-Informed Letter" invoking export-control emerging-tech authority and EAR §744.22(b) "military-intelligence end-use," ordering that exports of Mythos and Fable 5 — including transfers to foreign persons — require a license; this is the first time frontier AI has been named as a controlled dual-use item (law firm Mayer Brown called it "novel and unprecedented"); exemptions for trusted partners followed June 26 and the controls lifted July 1, restoring access to a US-government-approved org list (AWS, Apple, Cisco, CrowdStrike, Microsoft, Nvidia, Palo Alto…) — control, exemption, lift, three times in one month; (3) the moat changed material — from "technical lead" to "policy-volatility exposure": binding your interface to a single frontier vendor means your users' access sits inside another government's administrative discretion, and one letter can cut off your foreign users (ties to ENTRY 76's cutoff of foreign access); China is not a mirror — don't say "reciprocal control": the US governs the export of offensive capability while China governs content and data sovereignty (generative-AI filing, critical-infrastructure localization, government-classified on-prem) — different risk dimensions, but the same backdrop pulls the model entity into national-security governance and boosts China's open-weight "downloadable, runs-locally" narrative (DeepSeek V4 went GA open-source July 20, Opus-class, ~1/7 the price, Ascend-adapted); the sharpest primary-market signal is RealAI closing a several-hundred-million-yuan Series B on July 21 for a "secure trusted LLM system" (desensitization + secure routing so frontier capability comes in while business secrets don't leak out), serving government and state-owned enterprises. Rollback: calling "licensed release the industry default" overstates it — the statutory export license so far applies case-by-case to Anthropic alone, while OpenAI and Google are voluntarily coordinating with government; the US has published no formal capability-gating threshold, so "more capable → gated" is inference, not written rule; and GPT-5.6's July was a release, not a tightening.
Falsifier: if within a year the US writes a capability threshold into a formal ECCN rule applying to all frontier models, "case-by-case" upgrades to "institution."