"Limited preview" sounds harmless. In frontier AI, it can mean exclusive early access to the next layer of economic infrastructure — and when the government helps decide who gets in, the launch window becomes a government-shaped competitive moat.
A limited frontier-AI release is not just a safety measure. When the government influences which companies receive early access to models like OpenAI GPT-5.6 or Anthropic Fable 5 / Mythos 5, it may also be deciding which companies get to build, test, market, sell, raise capital, and scale with the next generation of AI before everyone else. That advantage tends to flow to large incumbents, Fortune 500 buyers, defense contractors, cloud partners, and politically connected institutions — while startups and smaller companies are left waiting.
The AI launch window
Early access to a frontier model is not an ordinary software beta. In frontier AI, a few weeks or months of advance access can translate into substantial commercial advantage: integrated products, signed enterprise contracts, captured workflow data, public credibility, and investor momentum. That is the AI launch window — and right now, the U.S. government is helping decide who steps through it first.
OpenAI has now launched GPT-5.6 Sol, Terra, and Lunain a limited preview.[1] During the preview, the models are available through the OpenAI API and Codex only to a limited group of selected trusted partners and organizations. GPT-5.6 is not available in ChatGPT during the preview, individual users cannot apply, and there is no public waitlist. OpenAI has stated that the limited preview was requested by and coordinated with the U.S. government, and that broader availability is planned "in the coming weeks" without a firm general-availability date.[3]
That is the public-facing product story.
The legal and competitive story is more complicated.
OpenAI says it previewed GPT-5.6's plans and capabilities to the U.S. government before launch. At the government's request, OpenAI is starting with a limited preview for trusted partners whose participation has been shared with the government.[3] OpenAI has also said it does not believe this kind of government access process should become the long-term default, because it keeps advanced tools away from users, developers, enterprises, cyber defenders, and global partners who need them.
OpenAI is cooperating, but it is also objecting. It is complying, but it is also warning that a temporary safety process could harden into a permanent access-control regime — one that doubles as a market-allocation regime.
Why GPT-5.6 triggered government attention
OpenAI's own materials make clear why the government cares. GPT-5.6 Sol is described as OpenAI's strongest model yet, with improvements in coding, biology workflows, and cybersecurity. OpenAI says GPT-5.6 Sol is its most capable cybersecurity model and shifts the frontier for long-horizon security tasks, including vulnerability research and exploitation.[1]
OpenAI's GPT-5.6 system card treats Sol, Terra, and Luna as "High" capability in both cybersecurity and biological/chemical risk under OpenAI's Preparedness Framework, while stating that the models do not reach the highest "Critical" cybersecurity level. The system card says the models can find vulnerabilities and pieces of exploits but were not able, in testing, to carry out autonomous end-to-end attacks against hardened targets.[2]
That is the dual-use problem in a single paragraph. The same model that can help defenders patch hospitals, banks, utilities, and government systems can also help bad actors discover weak points faster. OpenAI emphasizes that GPT-5.6 is better at finding and fixing vulnerabilities than reliably executing real attacks, and that broad access can produce safety benefits for defenders.
So the government's concern is not imaginary. But the legal question remains: concern alone is not authority.
Under what authority is the government acting?
The central authority is Executive Order No. 14,409, Promoting Advanced Artificial Intelligence Innovation and Security, 91 Fed. Reg. 34,565 (June 5, 2026).[4]
The order directs federal agencies to create a classified benchmarking process for advanced cyber capabilities and to determine when an AI model should be designated a "covered frontier model." The order assigns that determination to the Director of NSA, in consultation with the National Cyber Director, the Assistant to the President for Science and Technology, CISA, and other defense representatives.
The order then directs agencies to design a voluntary framework with AI developers. Under that framework, developers may engage with the government to determine whether a model is a covered frontier model, provide federal access to covered frontier models for up to 30 days before release to trusted partners, and collaborate with the government to select trusted partners for early access.
The order does not authorize mandatory governmental licensing, preclearance, or permitting for the development, publication, release, or distribution of new AI models, including frontier models.
That disclaimer is the legal fulcrum.
If the government is merely asking OpenAI to participate in a short, voluntary, pre-release safety review, the arrangement is probably on safer ground. If the government is actually deciding customer-by-customer who may access the model, or if OpenAI reasonably believes refusal would trigger retaliation, procurement consequences, export-control action, or other penalties, the arrangement starts to look less voluntary and more like informal licensing.
OpenAI versus Anthropic: request versus directive
The OpenAI situation appears to be a government-requested limited release under the voluntary framework. OpenAI has publicly described the restriction as occurring at the government's request, not as the result of a published binding legal order.
Anthropic is different.
Anthropic says the U.S. government issued an export-control directive requiring the company to suspend access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign-national Anthropic employees. Anthropic said the practical effect was that it had to disable those models for all customers to ensure compliance.[5]
That is not voluntary compliance. That is legal compliance with a government directive.
The government has since partially loosened restrictions on Mythos 5 for some trusted U.S. organizations, while restrictions remain for others and Fable 5's general release remains unresolved. Reuters reported that the government allowed Mythos 5 to be redeployed to more than 100 trusted U.S. organizations and that Anthropic is continuing to work with the government to expand Mythos access and make Fable 5 available for general use again.[6]
Government-requested limited preview
- ›GPT-5.6 Sol, Terra, Luna in limited preview to selected trusted partners
- ›Individual users cannot apply; no public waitlist
- ›OpenAI says the limited preview was requested by and coordinated with the U.S. government
Export-control directive
- ›Directive suspended Fable 5 / Mythos 5 access for foreign nationals
- ›Practical effect: disabled for all customers to ensure compliance
- ›Later partial access restored only for selected trusted U.S. organizations
The legal danger is that the soft gate may become the hard gate without Congress ever creating a formal frontier-AI licensing statute. The commercial danger is that, hard or soft, the gate decides who gets to compete in the next AI cycle.
Five advantages of advance access
Companies admitted to a frontier-model launch window gain at least five compounding advantages over competitors waiting on general availability:
Product-development advantage
Early-access companies build, test, and iterate against the next-generation model while competitors are still working against the prior frontier.
Customer-lock-in advantage
First movers sign enterprise contracts, embed workflows, and capture switching costs before smaller competitors can demo.
Data and implementation-learning advantage
Real workflow data, feedback loops, fine-tuning, and deployment know-how accumulate inside incumbents weeks or months ahead of the field.
Credibility and trusted-partner signaling
Being named a trusted partner functions as a government-endorsed quality signal that smaller competitors cannot replicate.
Capital, procurement, and enterprise-sales advantage
Investors, federal procurement officers, and Fortune 500 buyers prefer vendors already operating on the frontier model — compounding the moat.
None of these advantages depend on the early-access company being better. They depend on being earlier. Over a few launch cycles, "earlier" becomes "dominant."
What this means for startups and smaller companies
Smaller companies and startups are structurally disadvantaged in any government-shaped access process. They usually lack the government relationships, in-house compliance teams, lobbying presence, national-security contacts, procurement history, federal contracting experience, and enterprise distribution channels that large incumbents take for granted.
When trusted-partner status is shaped — even informally — by who already has those relationships, frontier-model access tracks existing market power instead of safety merit. Startups are then asked to compete against incumbents that already had weeks or months to integrate the new model, train staff, sign customers, and tell investors they were on the frontier first.
That is not a level playing field. That is a government-shaped launch window that quietly decides which companies get to write the next chapter of the AI economy.
The export-control theory
For Anthropic, the reported legal mechanism is export control. The likely statutory and regulatory sources are the Export Control Reform Act of 2018, 50 U.S.C. §§ 4801–4852, and the Export Administration Regulations, 15 C.F.R. pts. 730–774.
Under the EAR, releasing controlled technology or source code to a foreign person inside the United States can be treated as a "deemed export." See 15 C.F.R. § 734.13(b). A "release" can include visual inspection or oral or written exchanges of controlled technology or source code. See 15 C.F.R. § 734.15. BIS also uses end-use and end-user controls under Part 744, and BIS guidance states that it may inform a person that a license is required for a specific export, reexport, or in-country transfer because of unacceptable diversion or end-use risk.
But the authority is contested when applied to hosted AI model access.
CSIS reports that Commerce's Anthropic letter apparently relied on ECRA emerging-technology authority and EAR military-intelligence controls, but CSIS also notes that no full regulatory framework has been implemented for that statutory authority and that it is unclear whether remote access to a hosted model is a "release" of software or technology under the EAR.[8]
The Harvard Law Review reached a similar issue: when a user sends a prompt to a model running on Anthropic's servers and receives a response, it is unsettled whether the provider has exported the model, the model weights, or controlled technology at all. The user typically does not receive the model weights, architecture, or source code; the user receives an output.[9]
Is access to intelligence an export of technology, or is it a service?
If it is a service, export-control authority may be a strained fit. If it is technology access, Commerce has a stronger argument. If the government believes frontier models require a new category of access control, Congress should legislate it rather than forcing old export-control tools to do work they were not clearly designed to do.
Is OpenAI's compliance voluntary?
Legally, based on public reporting and OpenAI's statements, OpenAI's GPT-5.6 limited preview is best described as voluntary compliance with a government request — not compliance with a known binding export-control order.
But practical voluntariness is more complicated.
A company can "voluntarily" comply because it wants to preserve relationships with federal agencies, avoid an Anthropic-style directive, protect procurement opportunities, avoid being characterized as reckless, reduce regulatory scrutiny, or keep a path open for broad release. That kind of compliance may be rational, but it is not the same as free and unpressured choice.
The constitutional line is familiar. The government may persuade, warn, coordinate, and request cooperation. But it may not use threats, regulatory leverage, or informal pressure to accomplish indirectly what it could not lawfully command directly. See Bantam Books, Inc. v. Sullivan, 372 U.S. 58 (1963); National Rifle Ass'n of America v. Vullo, 602 U.S. 175 (2024).
That doctrine matters here. If the government says, "Please participate in a 30-day safety review," that is one thing. If the government says, "Limit access to these customers or face export controls, procurement consequences, criminal scrutiny, public condemnation, or national-security designation," that is something else.
The first is cooperation. The second may be coercion.
The Youngstown problem
The major constitutional question is whether the Executive is creating a licensing regime without Congress.
Under Youngstown Sheet & Tube Co. v. Sawyer, 343 U.S. 579 (1952), presidential authority is strongest when supported by Congress, uncertain when Congress has not spoken, and weakest when the Executive acts contrary to congressional limits.
Executive Order 14,409 tries to avoid the Youngstown problem by saying the framework is voluntary and does not create mandatory licensing or preclearance. That helps. But if the practical reality is that frontier AI companies need an executive "green light" before broad model release, courts and Congress may eventually treat the process as a de facto licensing regime.
That would raise major-questions concerns. Under West Virginia v. EPA, 597 U.S. 697 (2022), agencies need clear congressional authorization for actions of vast economic and political significance. After Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), courts will not automatically defer to an agency's interpretation of ambiguous statutory authority.
Frontier AI model release is economically and politically significant. If the government wants binding pre-release approval authority, it should obtain clear statutory authority.
The APA problem
If government involvement becomes final agency action, the Administrative Procedure Act comes into play.
Under Bennett v. Spear, 520 U.S. 154 (1997), reviewable final agency action generally requires the consummation of agency decision-making and legal consequences or practical effects. A mere request to OpenAI may not qualify. A written directive, license requirement, denied export authorization, or government-approved customer list may.
If agency action is reviewable, it must be reasoned and non-arbitrary. See Motor Vehicle Mfrs. Ass'n v. State Farm Mut. Auto. Ins. Co., 463 U.S. 29 (1983). That means the government should have evidence, standards, a record, and a rational connection between the risk and the restriction.
"Trusted partner" cannot mean "companies the government likes." It must mean "entities that satisfy published, objective, security-based criteria."
The practical risk: access control becomes market control
The most dangerous part of the OpenAI and Anthropic episodes is not the safety review. Safety review is defensible.
The dangerous part is customer selection.
Reuters reported criticism that no one knows how approved companies are picked or why others are excluded, and reported Sam Altman's concern that extensive safety testing may be reasonable but government customer selection is not.[7]
That is the issue LegalTek.ai should frame sharply: a temporary review of dangerous capability is not the same thing as a permanent government role in rationing access to intelligence infrastructure.
The more AI becomes infrastructure for cybersecurity, legal services, medicine, education, finance, government, logistics, and science, the more access decisions become market-structure decisions. If the government approves large incumbents first and startups later, that is not neutral. If approved users get frontier capability weeks or months earlier than others, that advantage can compound.
Antitrust and platform-access analysis
Frontier AI models are becoming critical inputs for legal tech, cybersecurity, medicine, finance, education, logistics, government services, and enterprise automation. When a small set of approved partners receives early, exclusive access to that input, the result looks less like a beta program and more like a platform-access problem.
Antitrust law has long recognized that control over a critical input can constrain competition downstream. The essential-facilities thread runs from United States v. Terminal Railroad Ass'n of St. Louis, 224 U.S. 383 (1912), through Associated Press v. United States, 326 U.S. 1 (1945), into modern platform cases such as United States v. Microsoft Corp., 253 F.3d 34 (D.C. Cir. 2001). When private actors control access to infrastructure that competitors need to reach customers, courts have policed that control carefully.
The analogy is imperfect — frontier-model providers are not common carriers, and AI models are not railroad terminals or wire services. But the structural concern is similar: if government-shaped access decides which downstream companies can build on the newest model, the government has effectively allocated a critical input. And under N.C. State Bd. of Dental Examiners v. FTC, 574 U.S. 494 (2015), a process that mixes government authority with incumbent private interests does not earn automatic antitrust immunity.
Safety testing vs commercial advantage
The right line is not between "any restriction" and "no restriction." It is between safety-testing access and commercial-development access.
Narrow early access for red-teaming, national-security testing, independent auditing, and critical-infrastructure defense may be justified — and in some cases necessary — before a frontier model reaches the public. That is a legitimate government interest and a legitimate use of voluntary cooperation.
But once early access permits product development, customer acquisition, marketing, investor signaling, procurement advantage, or commercial deployment, the analysis changes. At that point, access criteria must be neutral, objective, published, and realistically available to qualified smaller companies. Otherwise the safety frame is doing market-allocation work it was never meant to do.
The LegalTek.ai PRIME Protocol
To decide whether a frontier-model release process is truly voluntary, use the PRIME Protocol.
Published authority
The government must identify the legal source for its involvement — executive order, export-control law, procurement authority, national-security authority, or voluntary guidance.
Record of request
Any government request affecting release timing, customer access, model availability, or trusted-partner selection should be documented. Informal pressure creates legal risk.
Independent criteria
Access decisions must be based on objective safety and security criteria — not politics, incumbency, agency preference, personal relationships, procurement leverage, or private-market influence.
Meaningful choice
Participation is not truly voluntary if refusal would likely trigger retaliation, procurement consequences, export-control escalation, national-security designation, loss of unrelated government benefits, or public punishment.
Expiration
Any limited release, pause, or restricted-access period must have a defined end date, renewal standard, and path to broader availability.
Applied to OpenAI, the current facts are mixed. There is published authority in the form of Executive Order 14,409, and OpenAI says the measure is temporary. But the trusted-partner process remains opaque, and the practical voluntariness depends on what was said privately between OpenAI and federal officials.
The LegalTek.ai NOVA Access Rule
PRIME asks whether the government's process is truly voluntary. The NOVA Access Rule asks the next question: whether early access is fair and open to qualified smaller players — startups, universities, nonprofits, researchers, and non-incumbents.
Neutral eligibility
Access criteria must apply equally to large companies, startups, universities, nonprofits, researchers, and smaller businesses.
Objective safeguards
"Trusted" must mean the applicant satisfies defined security requirements — identity verification, logging, monitoring, abuse controls, incident response, data handling, and use-case limits.
Visible process
The government and model provider should publish redacted criteria, timelines, approval and denial categories, aggregate statistics, and general mitigation requirements.
Appeal and cure
Denied applicants should receive a reason, a chance to fix deficiencies, and an expedited review path where delay would cause serious competitive or security harm.
A frontier-AI access regime that cannot survive NOVA is not a safety program. It is a market-allocation program wearing a safety label.
The VECTOR Test for restrictions
Before the government restricts or shapes access to a frontier model, it should satisfy the VECTOR Test.
Valid authority
What statute, executive order, regulation, procurement rule, export-control authority, or other lawful basis supports the action?
Evidence of elevated risk
What specific capability makes this model materially riskier than prior public models?
Competitive-impact analysis
Who benefits from early access, who is excluded, and how will the government prevent access from becoming a market moat?
Tailored restriction
Why is the restriction narrower and better than available alternatives?
Oversight record
What written record supports the decision?
Review path
Can the developer or excluded applicant challenge, cure, reapply, or seek expedited review?
OpenAI's limited preview appears stronger on evidence than process. GPT-5.6 has documented cyber and bio/chemical risk characteristics. But the public still lacks the access criteria, approval process, denial process, and government role in customer selection.
Anthropic's export-control directive appears stronger on formal compulsion than transparency. It had legal teeth, but the publicly available record raises serious questions about fit, scope, and whether hosted model access is properly treated as an export.
How PRIME, NOVA, and VECTOR work together
The three LegalTek.ai frameworks answer three different questions about a government-shaped frontier-AI release:
- PRIME — Is the government's "voluntary" process actually voluntary, or is it informal coercion dressed up as cooperation?
- NOVA — Is early access fair and open to qualified smaller players, or is "trusted partner" a euphemism for "already powerful"?
- VECTOR — Is any resulting restriction legally justified, evidence-based, tailored, and reviewable?
A frontier-AI release process that fails any of the three should not be defended as ordinary safety practice. PRIME guards the constitutional line. NOVA guards the competitive line. VECTOR guards the administrative-law line. Together they describe the minimum a defensible launch-window regime should look like.
FAQ: voluntary, coercive, or market-distorting?
Common questions about how the PRIME Protocol, NOVA Access Rule, and VECTOR Test distinguish lawful safety cooperation from informal coercion and government-shaped market allocation.
How do PRIME, NOVA, and VECTOR decide whether government involvement is truly voluntary?
PRIME runs the voluntariness test first. A frontier-model release process is genuinely voluntary only when the government identifies a published legal authority, documents its requests in a written record, applies independent and objective criteria, leaves the developer with a meaningful choice to refuse without retaliation, and attaches a clear expiration date to any limited-access period. If any one of those is missing, the process is no longer voluntary in any meaningful constitutional sense — it is voluntary on paper and coercive in practice.
When does government involvement cross the line from cooperation into coercion?
Coercion appears when refusal would predictably trigger retaliation, procurement consequences, export-control escalation, national-security designation, regulatory scrutiny, or loss of unrelated federal benefits. Under Bantam Books, Inc. v. Sullivan, 372 U.S. 58 (1963), and National Rifle Ass'n of America v. Vullo, 602 U.S. 175 (2024), the government may persuade and request, but it may not use informal pressure to accomplish indirectly what it could not lawfully command directly. PRIME's 'Meaningful choice' factor is designed to surface exactly that line.
When does a limited frontier-AI release become market-distorting?
NOVA flags market distortion the moment 'trusted partner' status stops tracking neutral safety criteria and starts tracking existing market power. If access criteria are not neutral across large companies, startups, universities, nonprofits, and researchers — or if the process lacks visible criteria, published timelines, and a real appeal-and-cure path — early access stops being a safety program and becomes a government-shaped allocation of competitive advantage to incumbents.
How does VECTOR evaluate a specific access restriction?
VECTOR is the legal-justification test for any restriction the government imposes or encourages: Valid authority (what statute, executive order, or regulation supports it), Evidence of elevated risk (what this model can do that prior models could not), Competitive-impact analysis (who benefits, who is excluded, and how the moat is prevented), Tailored restriction (why a narrower alternative would not work), Oversight record (the written basis for the decision), and Review path (the ability of developers or excluded applicants to challenge, cure, or reapply). A restriction that cannot satisfy VECTOR is vulnerable under West Virginia v. EPA, 597 U.S. 697 (2022), and State Farm, 463 U.S. 29 (1983).
Can a process be lawful under PRIME and VECTOR but still fail NOVA?
Yes — and that is the most important case. A limited preview can have valid authority, evidence of elevated risk, and a reasoned record, yet still allocate early access only to large incumbents, defense contractors, and politically connected institutions. PRIME and VECTOR can be satisfied while NOVA fails. When that happens, the safety frame is doing market-allocation work it was never designed to do, and the launch window itself becomes the harm.
What is the practical takeaway for founders, GCs, and policymakers?
Treat any government-shaped frontier-AI release as a three-part test. Run PRIME to see whether the 'voluntary' label survives scrutiny. Run NOVA to see whether the access process is realistically open to qualified smaller players. Run VECTOR to see whether any resulting restriction is legally defensible. If a program cannot pass all three, it should be redesigned before it hardens into informal licensing or a permanent incumbent moat.
Bottom Line
Safety testing may be limited. Commercial advantage must be fair. The government may regulate genuinely dangerous frontier-AI capabilities, but it should not allow "trusted partner" access to harden into a monopoly moat for incumbents, defense contractors, cloud partners, and politically connected institutions.
A voluntary safety process should not become informal licensing. A national-security review should not become government-shaped market allocation. And a temporary launch window should not become a permanent advantage for whoever was already on the inside.
Voluntary safety review is permissible.
Secret access control is dangerous.
Binding restriction requires clear authority.
Customer selection requires neutral, published process.
Commercial advantage must not be government-allocated.
Use PRIME to test voluntariness. Use NOVA to test fairness. Use VECTOR to test restrictions. And never let "trusted partner" become a euphemism for "government-approved insider."
References & citations
Numbered footnotes cover factual claims about OpenAI, Anthropic, government process, executive orders, and reporting. In-text case and statute citations link directly to Justia, Cornell LII, or the eCFR.
- [1]OpenAI, Introducing GPT-5.6 (Sol, Terra, Luna) — limited preview announcement and system card overview, openai.com (2026). https://openai.com/index/
- [2]OpenAI, GPT-5.6 System Card — Preparedness Framework categorizations for cybersecurity and biological/chemical risk, openai.com (2026). https://openai.com/safety/
- [3]OpenAI public statements describing the GPT-5.6 limited preview as requested by and coordinated with the U.S. government, openai.com (2026). https://openai.com/global-affairs/
- [4]Exec. Order No. 14,409, Promoting Advanced Artificial Intelligence Innovation and Security, 91 Fed. Reg. 34,565 (June 5, 2026). https://www.federalregister.gov/presidential-documents/executive-orders
- [5]Anthropic statement describing a U.S. government export-control directive suspending Fable 5 / Mythos 5 access for foreign nationals, anthropic.com (2026). https://www.anthropic.com/news
- [6]Reuters, reporting on partial restoration of Mythos 5 access to trusted U.S. organizations and ongoing Fable 5 negotiations, reuters.com (2026). https://www.reuters.com/technology/artificial-intelligence/
- [7]Reuters, reporting on criticism that approved-partner selection criteria are not publicly disclosed, and Sam Altman comments on government customer selection, reuters.com (2026). https://www.reuters.com/technology/artificial-intelligence/
- [8]Center for Strategic & International Studies, analysis of the Commerce/BIS letter to Anthropic, ECRA emerging-technology authority, and EAR military-intelligence controls, csis.org (2026). https://www.csis.org/analysis
- [9]Harvard Law Review, note on whether remote access to a hosted AI model constitutes a 'release' of technology or source code under the Export Administration Regulations, harvardlawreview.org (2026). https://harvardlawreview.org/
- [10]Export Control Reform Act of 2018, 50 U.S.C. §§ 4801–4852. https://www.law.cornell.edu/uscode/text/50/chapter-58
- [11]Export Administration Regulations, 15 C.F.R. pts. 730–774 (definitions of 'release' and 'deemed export' at §§ 734.13, 734.15). https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-C
Disclaimer: This article is for general informational and educational purposes and does not constitute legal advice. No attorney-client relationship is created by reading this material. LegalTek.ai is a technology company, not a law firm.