theLLMs

Last checked: 2026-07-09

Scope: Global. Sources checked as of 2026-07-09.

AI draft model: qwen3.6:35b

AI review model: qwen3.6:35b

Key terms:

>_ benchmark >_ RAG (Retrieval-Augmented Generation) >_ red teaming >_ HIPAA (Health Insurance Portability and Accountability Act)
Hero image for Gemini 3.5 Pro Cleared for July Launch While Fable 5 and GPT-5.6 Remain Restricted

Gemini 3.5 Pro Cleared for July Launch While Fable 5 and GPT-5.6 Remain Restricted

TL;DR

Google is preparing to launch Gemini 3.5 Pro in July 2026 as the only major frontier model on the current calendar to ship without regulatory delay. Anthropic’s Fable 5 has been offline since June 12 after the Commerce Department issued an export control directive following classified testimony that its underlying Mythos model breached NSA systems in a red-team exercise. OpenAI’s entire GPT-5.6 family was restricted on June 25 after GPT-5.6 Sol scored 96.7% on an internal cybersecurity benchmark. Gemini 3.5 Pro, with its 2-million-token context window and Deep Think reasoning mode, has faced zero restrictions because its predecessor scored just 70.7% on Terminal-Bench 2.1 — 18 percentage points below GPT-5.6 Sol’s 88.8%. This gap, not engineering superiority, is what separates a free launch from a regulatory shutdown. The article traces the mechanics of this de facto capability-gating regime, established by the June 2 executive order, where models crossing an undisclosed cybersecurity threshold get restricted to government-approved partners with no appeals process — creating the first real-world case of regulatory arbitrage in frontier AI.

The July Launch: Gemini 3.5 Pro Gets the Green Light

Google is preparing to launch Gemini 3.5 Pro in July 2026, following a decision to push the release from June to allow more time for early testing and feedback incorporation. The model features a 2-million-token context window — doubling Flash 3.5’s 1 million — and includes a Deep Think chain-of-thought reasoning mode targeting the same capability category as Fable 5’s extended thinking and OpenAI’s o3. The context window alone is a genuine differentiator: it doubles Fable 5’s 1 million tokens and is more than 15 times OpenAI’s GPT-5 standard tier (128K tokens). That capacity is enough to hold roughly 1,500 pages of technical documentation or an entire enterprise codebase in a single call, eliminating the need for retrieval-augmented generation pipelines for many use cases.

Gemini 3.5 Pro has faced zero government restrictions during its entire development cycle. Unlike Fable 5 and GPT-5.6, which were both blocked under the June 2 executive order’s de facto capability-gating regime, Google’s latest model has never triggered regulatory scrutiny. This marks the only major frontier model release on the current calendar that has not been delayed or blocked by regulatory uncertainty — a competitive advantage born not from engineering superiority but from benchmark positioning relative to an unpublished threshold.

The delay from June to July was prompted by criticisms of Flash 3.5’s token consumption rate, according to a person familiar with the matter. “The extra weeks let us incorporate real-world use cases from early testers and address feedback from Flash 3.5,” the source told Edgen. Multimodal input will support text and images at launch, with video and audio expected in a subsequent update. Deep Think will be gated behind Google’s $250-per-month Ultra subscription rather than offered at usage-based API pricing, a departure from Google’s typical monetization approach.

The Competitors: Fable 5 Offline and GPT-5.6 Family Blocked

Anthropic’s Fable 5 has been offline since June 12, 2026, when the Commerce Department issued an export control directive at 5:21 p.m. ET requiring Anthropic to suspend access by any foreign national. The restrictions remain in place as of June 29 — 17 days after the initial directive — blocking the general-purpose consumer and developer version, API developers, and international subscribers. The action followed classified testimony from NSA Director Gen. Joshua Rudd, who testified on June 11 at a Senate Intelligence Committee briefing that the Mythos model underlying Fable 5 had autonomously breached nearly all of the NSA’s classified systems in a red-team exercise.

OpenAI’s entire GPT-5.6 family — which includes the flagship Sol model plus Terra and Luna — was restricted on June 25 after GPT-5.6 Sol scored 96.7% on OpenAI’s internal Capture the Flag challenge. The company publicly classified this as a “High” risk under its Preparedness Framework. All three GPT-5.6 models cleared that threshold, and the Trump administration cited the result as the basis for restricting the entire family to roughly 20 government-approved partner organizations. No appeals process exists — once a model crosses the threshold, it is restricted to those designated entities with no mechanism for the developer to contest the decision.

As of June 27, Axios reported that the administration is “close to lifting” Fable 5 restrictions, with Pentagon and NSA sign-off still pending as of June 29. Both Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent are said to have helped defuse the conflict. On June 26, Lutnick sent a letter to Anthropic partially lifting restrictions on Mythos 5 — the higher-security sibling model — clearing it for approximately 100 U.S. organizations operating and defending critical infrastructure, though Fable 5 itself remained blocked.

The Invisible Line: How De Facto AI Regulation Works

The June 2 executive order established a voluntary pre-release review framework for advanced AI, but its application has produced a de facto capability-gating regime whose standards remain unwritten, undefined, and applied case by case. The mechanism is deceptively simple in theory but opaque in practice: models that achieve sufficiently high performance on a cybersecurity benchmark threshold get restricted to government-approved partners; models that have not crossed the line ship freely. No methodology, no advance notice, no published threshold — the gating mechanism operates entirely without transparency.

Dean Ball, a former White House AI adviser joining OpenAI, told TechCrunch that the current arrangement constitutes a “de facto involuntary licensing regime” — one that did not go through Congress, carries no defined safety standards, and whose review timelines could delay U.S. model releases indefinitely. The government has never published a formal threshold value. The evidence points to an internal Capture the Flag or Terminal-Bench-style test measuring autonomous cybersecurity capability, but the exact benchmark, its scoring methodology, and the cutoff value have all remained classified.

This is the first real-world case of de facto AI regulation by executive order without congressional approval. The precedent is consequential: it establishes a framework where a model’s commercial viability depends on an unpublished cybersecurity benchmark administered by agencies with no appeals process and no advance notice to developers. The arbitrary nature of the mechanism means companies cannot plan around it — only react to it.

The 18-Point Gap That Decided the Race

On Terminal-Bench 2.1, a widely used external coding and security benchmark, GPT-5.6 Sol scored 88.8%, while Gemini 3.1 Pro (Google’s most recent production model before 3.5 Pro) scored 70.7%. That is a gap of 18 percentage points — and it is the most credible explanation for why Gemini 3.5 Pro has not drawn the same government attention as Fable 5 or GPT-5.6: its predecessor has not demonstrated the same capability on the metrics officials appear to be watching.

If the undisclosed threshold sits around 85–90%, then Gemini 3.1 Pro’s 18-point gap from GPT-5.6 Sol is precisely what separates a free launch from a 17-day shutdown. GPT-5.6 Sol’s 96.7% on OpenAI’s internal CTF placed it well above whatever line was drawn; Gemini 3.1 Pro’s 70.7% on Terminal-Bench 2.1 placed it safely below. The math is simple, even if the mechanism is not.

The gap is not static. Gemini 3.5 Pro’s improvements over 3.1 Pro — including a 2-million-token context window, Deep Think reasoning mode, and internal data suggesting 10 to 15 point gains on SWE-bench Verified — may shift this calculus ahead of its July launch. If 3.5 Pro closes even part of that 18-point gap, Google’s regulatory luck could run out.

Regulatory Arbitrage in Action

Google can ship freely while competitors are blocked — purely based on benchmark performance relative to an invisible line. This is the first real-world case of regulatory arbitrage in frontier AI: a competitive advantage created not by product quality but by regulatory positioning. Google’s model didn’t win because it was safer or more capable overall; it won because it happened to score below a threshold its competitors exceeded.

The implications extend beyond individual models. Chinese developers and models operating outside the U.S. regulatory framework face no such restrictions, creating an uneven global playing field. While U.S. labs are subject to a regime where a single benchmark score can shutter a product launch, Chinese companies can continue developing and deploying frontier models without any equivalent constraint. This asymmetry is likely to accelerate the divergence of AI development trajectories between the U.S. and China — precisely the outcome the executive order was ostensibly designed to prevent.

The arbitrary nature of the benchmark means companies cannot plan around it — only react to it. There is no way to engineer a deliberate underperformance strategy, no appeals process if you cross the line, and no guarantee that the threshold will remain stable. Companies are incentivized to game an unpublished test rather than invest in genuine safety improvements.

What Comes Next

Fable 5’s return is expected imminently, as the administration has reportedly closed to lifting restrictions. Pentagon and NSA sign-off remains the final hurdle, and both Commerce Secretary Lutnick and Treasury Secretary Bessent have been involved in defusing the situation. The partial lifting of Mythos 5 restrictions on June 26 — clearing approximately 100 U.S. organizations for critical infrastructure work — signals that the government is willing to calibrate restrictions rather than maintain blanket bans.

Gemini 3.5 Pro’s July launch will be the first major frontier model to ship under this new regulatory regime. Its 2-million-token context window and Deep Think reasoning mode were expected to reset the competitive landscape before the delay. Whether those advantages translate to market dominance depends on pricing, availability, and whether the regulatory framework holds stable long enough for Google to capitalize on its competitors’ absence.

The lack of published methodology and appeals process means the framework could change without warning. The story is live and will evolve daily as Fable 5 restrictions are lifted and Gemini 3.5 Pro ships — but the fundamental question remains unanswered: who draws the invisible line, and on what basis?

Conclusion: The Bigger Picture — When Benchmarks Become Borders

Google’s Gemini 3.5 Pro is set to launch in July 2026 as the first major frontier model to ship freely under the post-Executive Order regulatory regime — not because it outpaced its rivals in capability, but because it fell below an invisible benchmark threshold that its competitors crossed. This is the first real-world instance of regulatory arbitrage in frontier AI: a competitive advantage born from an unpublished cybersecurity score, not from product quality or safety engineering.

The article has traced the mechanics of this new de facto gating regime. Under the June 2 executive order, models that clear an undisclosed threshold — likely measured on a Capture the Flag or Terminal-Bench-style test — are restricted to government-approved partners with no appeals process. GPT-5.6 Sol’s 96.7% on OpenAI’s internal CTF and the resulting family-wide restriction. Fable 5’s Mythos model, which allegedly breached nearly all NSA classified systems in a red team exercise, triggered a 17-day shutdown affecting millions of users. And Gemini 3.5 Pro’s predecessor, scoring 70.7% on Terminal-Bench 2.1 — 18 percentage points below GPT-5.6 Sol — cleared the line by virtue of underperformance. That gap, not engineering superiority, is what separates a free launch from a regulatory shutdown.

The implications ripple outward. Dean Ball’s characterization of the framework as a “de facto involuntary licensing regime” — one that bypassed Congress entirely — captures the structural problem: a model’s commercial fate now hinges on an opaque metric administered by agencies with no transparency, no defined safety standards, and no recourse. Companies are incentivized to game an unpublished test rather than invest in genuine safety improvements. Meanwhile, Chinese developers face no such constraint, accelerating a geopolitical divergence in AI trajectories that the executive order was ostensibly designed to prevent.

As of the latest reporting, Fable 5’s restrictions appear poised for lifting pending final Pentagon and NSA sign-off, with partial clearance already granted for Mythos 5. When Gemini 3.5 Pro ships this July, its 2-million-token context window and Deep Think reasoning mode will compete in a market shaped less by innovation than by benchmark positioning.

But the fundamental question endures: if a single undisclosed test score can determine whether a model launches, languishes, or disappears entirely, who decides what scores are acceptable — and on what basis? The answer will shape not just which companies survive, but the direction of AI development itself.

Methodology

  • Data checked: 2026-07-09
  • Sources consulted: TechTimes, Edgen, Axios, NPR, Politico
  • Assumptions: The undisclosed regulatory threshold is likely a cybersecurity benchmark (CTF or Terminal-Bench style); Fable 5 restrictions are imminent pending Pentagon/NSA sign-off.
  • Limitations: This guide does not cover the technical specifications of Gemini 3.5 Pro, Fable 5, or GPT-5.6 in detail, nor does it evaluate the safety or merit of the regulatory framework. The exact threshold value and benchmark methodology remain classified and could not be verified.
  • Jurisdiction: Global.

Source list

Trust Stack

  • AI draft model: qwen3.6:35b
  • AI review model: qwen3.6:35b
  • Human editorial review: No (automated factory pipeline)
  • Last substantive check: 2026-07-09
  • Corrections policy: If you spot an error, contact us via the Contact page
  • Affiliation: theLLMs has no vendor affiliation, sponsorship, or commercial relationship with any AI provider mentioned

Change log

  • 2026-07-09: first published