Seventeen days after the U.S. government took Anthropic's Claude Fable 5 offline in the first forced shutdown of a commercially deployed frontier AI model, a resolution is finally approaching — and it arrives just as Google prepares to fill the vacuum with Gemini 3.5 Pro. Axios reported on June 27 that the Trump administration is close to lifting the restrictions on Fable 5, with insiders expecting access to return as soon as this week. Meanwhile, Gemini 3.5 Pro — already delayed from June to July — has never been restricted at all, leaving Google as the only major frontier AI lab set to release a new top-tier model without government interference.

That gap is not a coincidence. The same June 2 executive order that set up a voluntary pre-release review framework for advanced AI has produced a de facto capability-gating regime whose standards remain unwritten, undefined, and applied case by case. Models that clear a cybersecurity benchmark threshold the administration has never publicly specified get restricted. Models that haven't crossed that line — including, so far, Gemini 3.5 Pro — ship freely. The result is a competitive landscape reshaped more by Washington than by engineering.

How Restrictions Work: An Undefined Cybersecurity Threshold

The restrictions imposed on Fable 5 on June 12 and on OpenAI's GPT-5.6 on June 25 both trace to the same technical concern: that models achieving sufficiently high performance on cybersecurity benchmarks could, in the wrong hands, enable automated vulnerability discovery and exploitation at a scale beyond existing defenses. The government has never published a formal threshold. But the evidence points to one.

GPT-5.6 Sol — the flagship of OpenAI's newest model family, which also includes Terra and Luna — scored 96.7% on OpenAI's internal Capture the Flag challenge, a test designed to measure autonomous cybersecurity capability, crossing what the company publicly describes as its Preparedness Framework's "High" risk classification. All three GPT-5.6 models cleared that threshold, which the Trump administration cited as the basis for restricting the entire family to roughly 20 government-approved partner organizations. On Terminal-Bench 2.1, a widely used external coding and security benchmark, Sol scored 88.8%.

Gemini 3.1 Pro, Google's most recent production model before the upcoming Pro launch, scored 70.7% on Terminal-Bench 2.1 — a gap of more than 18 percentage points from Sol. That measured performance gap 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: it hasn't demonstrated the same capability on the metrics officials appear to be watching.

The government has not named any threshold. What exists instead is an informal, undocumented standard applied by executive agencies with no appeals process, no published methodology, and no advance notice to developers. Dean Ball, a former White House AI adviser who is joining OpenAI, told TechCrunch 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 while delivering advantage to Chinese AI developers operating outside the framework entirely.

Fable 5 on the Verge of Return After 17 Days Offline

The Fable 5 episode, which began when the Commerce Department issued an export control directive at 5:21 p.m. ET on June 12 requiring Anthropic to suspend access by any foreign national, is now entering what may be its final days. Axios reported on June 27 that the administration is close to lifting the 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, with one administration source telling Axios that Anthropic "has worked positively with the government" — a marked shift from Defense Secretary Pete Hegseth's March designation of the company as a "Supply-Chain Risk to National Security."

On June 26, Lutnick sent a letter to Anthropic's chief compute officer Tom Brown partially lifting restrictions on Mythos 5 — the higher-security sibling model — clearing it for approximately 100 U.S. organizations operating and defending critical infrastructure. The Fable 5 restrictions, which blocked the general-purpose consumer and developer version, remained in place. As of June 29, Fable 5 is still offline for general users, API developers, and international subscribers.

The government initially acted after Amazon researchers demonstrated a bypass technique to U.S. officials — a technique Anthropic publicly disputed as narrow, non-universal, and replicable on other publicly available models including OpenAI's GPT-5.5. The fuller picture from subsequent reporting was starker: NSA Director Gen. Joshua Rudd testified in a Senate Intelligence Committee briefing on June 11, the day before the directive arrived, that Mythos — the underlying model Fable 5 shares its weights with — had autonomously breached nearly all of the NSA's classified systems in a classified red-team exercise.

Gemini 3.5 Pro in July: Google's Opening

Gemini 3.5 Pro missed its June launch target and is now expected to reach general availability in July, according to a Business Insider report citing a source familiar with the matter. Google has attributed the delay to addressing token efficiency issues flagged by early testers and incorporating lessons from the Gemini 3.5 Flash rollout. A Google spokesperson declined to comment on the timeline.

The delay arrives alongside an unsettling personnel signal. In the week of June 21-27, four senior Gemini researchers announced they were leaving Google for Anthropic, following a broader pattern of talent movement away from the company's AI division. These are product and organizational challenges independent of the regulatory situation — but they arrive at a moment when the stakes of a timely July launch have sharpened considerably.

The model itself represents a meaningful step in context capacity. Gemini 3.5 Pro is expected to feature a 2-million-token context window — double the capacity of most frontier competitors and the largest in any production model — alongside a "Deep Think" reasoning mode available to $250-a-month Ultra plan subscribers. Gemini 3.5 Flash, which launched at Google I/O on May 19, already outperformed Gemini 3.1 Pro on most coding and agentic benchmarks at lower cost; Pro is positioned to close the remaining gaps on harder reasoning and long-context retrieval. Pricing is expected to track prior generations at approximately $15 per million input tokens and $60 per million output tokens, with a 10x premium for the Deep Think reasoning mode restricted to Ultra-tier subscribers.

For developers who built workflows around Fable 5's long-context capability before June 12, Gemini 3.5 Pro's 2-million-token window would directly address the gap left by the suspension. Whether Fable 5 returns before Gemini 3.5 Pro ships — and under what access conditions — will determine whether Google's window is a genuine competitive opening or a brief passing advantage.

What GPT-5.6 Users Are Still Waiting For

OpenAI's GPT-5.6 family launched on June 25 under an access restriction the company publicly resisted even as it complied. The launch is limited to a small group of trusted partners whose participation was shared with the government, with a broader rollout expected in coming weeks contingent on the review process. OpenAI stated in its announcement that it did not believe the government access process "should become the long-term default," calling it a threat to users, developers, enterprises, and cyber defenders.

The restriction came after OpenAI and the administration jointly assessed GPT-5.6 as reaching Mythos-level cybersecurity capability — a determination that triggered the same type of voluntary-but-effectively-mandatory review that preceded the forced shutdown of Fable 5. Unlike Anthropic, which fought the directive publicly and is still negotiating, OpenAI accepted the terms in advance, negotiating a controlled rollout rather than a post-launch suspension.

The government's reasoning relies on a framework Congress has never voted on and that the administration itself has not fully defined. As of June 29, the White House is still working with AI labs to establish what criteria will govern the process, which agencies have authority to act, and how long a review can last. The gap between the government's security concerns and its regulatory capacity to address them is where the industry now operates.

What Frontier AI Access Looks Like in July

For developers and enterprise customers evaluating options in the month ahead, the picture depends on timing across three open situations simultaneously.

Gemini 3.5 Pro is the only major upcoming frontier model release without a current government access restriction. Its 2M token context window is the largest in any production model and addresses the long-context gap left by Fable 5's suspension. Gemini 3.5 Flash, available now through the Gemini API, Google AI Studio, and Vertex AI, offers a bridge option that already outperforms Gemini 3.1 Pro on most coding and agentic benchmarks.

Fable 5 may return this week — but under conditions that remain unspecified. Anthropic's July 8 identity verification rollout appears to be the likely domestic access pathway, and whether the Annex A partner restrictions that currently apply to Mythos 5 will also cover Fable 5's general return has not been confirmed. Developers should not assume the Fable 5 that returns will have the same access model as the one that launched June 9.

GPT-5.6's broader rollout, expected in coming weeks from June 25, puts it on a trajectory for general availability in mid-to-late July — behind Gemini 3.5 Pro's targeted schedule.

The broader implication is structural. Frontier AI model availability is now a policy variable, not just a commercial one. The suspension of Fable 5 and the gating of GPT-5.6 have established a new risk category: a commercial AI service can be pulled globally, instantly, without prior statutory process and without an appeals mechanism, by a government directive that applies an undefined capability threshold no one has voted on and no developer can fully anticipate. Gemini 3.5 Pro is on the right side of that threshold today. Whether it remains there after its July launch, once government evaluators have a chance to measure the Pro tier's cybersecurity performance directly, is a question no one at Google has yet answered publicly.

Frequently Asked Questions

When will Gemini 3.5 Pro be released?

As of June 29, 2026, Gemini 3.5 Pro remains in limited preview for select Vertex AI enterprise customers. Google is targeting general availability in July 2026, based on a report from a source familiar with the matter. No specific release date has been announced, and Google has declined to comment on the timeline.

Why is Gemini 3.5 Pro not restricted by the government while Fable 5 and GPT-5.6 are?

The U.S. government has not published a formal capability threshold for its AI review process, but available benchmark data points to cybersecurity performance as the operative variable. GPT-5.6 Sol scored 96.7% on OpenAI's internal Capture the Flag evaluation and crossed its Preparedness Framework's "High" risk classification. Gemini 3.1 Pro, Google's most recent production model, scored 70.7% on Terminal-Bench 2.1 — over 18 percentage points below Sol. That performance gap is the evidence-supported explanation for why Gemini models have not triggered the government's informal review threshold.

Is Fable 5 coming back?

As of June 29, 2026, Fable 5 remains offline on Day 17 of its suspension. Axios reported on June 27 that the Trump administration is close to lifting the restriction, with insiders expecting access to return as soon as this week, pending Pentagon and NSA sign-off. A Commerce Department letter on June 26 partially restored Mythos 5 for approximately 100 U.S. critical infrastructure organizations — but the Fable 5 general-user restriction has not yet been lifted.

What does the government's AI review regime actually require?

The June 2 executive order asked advanced AI companies to voluntarily submit models for government review up to 30 days before release. As of June 29, 2026, no specific criteria have been published for what triggers a restriction, what benchmarks are evaluated, which agencies have authority to act, or how long a review can last. Dean Ball, a former White House AI adviser, has described the current arrangement as a de facto involuntary licensing regime — one that functions without statutory authorization, published standards, or any appeals mechanism.