Sam Altman is letting Howard Lutnick pick the customers for GPT-5.6.
That is what OpenAI’s Friday announcement amounts to, once you read past the language about “transition periods” and “broader availability.” OpenAI’s newest model — three variants under the GPT-5.6 umbrella, tuned for reasoning, everyday tasks, and heavier workloads — will reach users only after the Trump administration signs off on who is allowed to have it. This is, by any plain reading, a company handing a customer-veto to the executive branch. And OpenAI is doing it the same week the administration shut Anthropic’s most powerful models out of foreign hands entirely.
It is worth being precise about what the model is, because the public discussion of AI capability has the misleading habit of treating “the model” as a thing rather than a continually-tuned set of weights serving a continually-revised objective function. GPT-5.6 is, per OpenAI’s announcement, materially better at autonomous task completion and at biology and cybersecurity problems — the same domains the company says it wants cyber defenders and “global partners” to use the models to bolster. The most capable current-generation models are now the most capability-concentrated, and decisions about who gets them are not technical footnotes.
The approval mechanism is half-built and partly informal. The Commerce Department’s Center for AI Standards and Innovation — CAISI, the unit that already tests other models — has been reviewing GPT-5.6. A recent executive order gave cybersecurity and national-security officials a bigger role in that evaluation. Altman and Lutnick discussed the phased rollout on Wednesday. The customers who will initially receive GPT-5.6 are “approved by the Trump administration,” per the Wall Street Journal’s reporting, not approved against any published standard OpenAI or anyone else can read. There is no list of criteria, no appeals process, no sunset. The model will become “generally available in the coming weeks,” OpenAI says, with the implicit understanding that “coming weeks” is a window the administration can extend whenever it chooses.
The Anthropic comparator sharpens the picture. Two of Anthropic’s newest models — Fable 5 and Mythos 5 — are now under a Trump administration ban on foreign access. Anthropic complied by halting all access, period, including for U.S. customers; the company had previously worked with the administration on a limited Mythos rollout. The latest Mythos models have been off for two weeks. Cybersecurity experts have, in the interim, urged the White House to reverse the restrictions on the grounds that defenders need the tools, not the other way around. The same logic applies, more sharply still, to GPT-5.6: the very capabilities — biology, cyber, autonomous task completion — that justify the administration gating access are the capabilities the administration’s own defenders say they need.
This is regulatory capture in its inverse form. The familiar version has an industry capturing its regulator. The current version has a regulator capturing the industry, by handing itself discretionary power to approve or withhold the most powerful commercial AI systems on national-security grounds nobody outside the executive branch gets to read. Cui bono: the administration gets a lever over the AI industry’s most valuable outputs, OpenAI gets to be the model that is currently approved, and the users, developers, enterprises, cyber defenders, and global partners OpenAI itself named as needing these tools lose access to them. Anthropic, for the moment, gets neither customers nor goodwill. The enterprise base that has built the last two years of AI-integration roadmaps on the assumption that frontier models will be reachable is now downstream of an opaque approval process with no published criteria, no appeals mechanism, and no statutory ceiling on how long the same office can keep the gate closed.
Cory Doctorow’s framework for platform decay — which he has termed “enshittification” — identifies four historical constraints on corporate extraction: competition, regulation, self-help, and labor. When a firm neutralizes competition by convincing the state to manage regulation on a discretionary, case-by-case basis, the remaining constraints evaporate. OpenAI has chosen to submit its neural networks to the Commerce Department, trusting the administration’s favor after securing a direct line to the Secretary. Co-founder and President Greg Brockman has donated to a Trump-aligned political-action committee. The difference between a functioning market and a state-licensed cartel is no longer just the coercion — it is the political alignment of the firms doing the negotiating.
OpenAI’s framing deserves close reading. The company warns that “this kind of government access process” should not “become the long-term default” because it “keeps the best tools from users, developers, enterprises, cyber defenders and global partners who need them.” That is a true sentence about a regime OpenAI is, in the same blog post, cooperating with. The structural explanation is not hard to see. OpenAI has, as a matter of public record, backed the administration’s broader posture on AI — staying ahead of China, accelerating the data-center buildout. The company is, in plain terms, cooperating with a discretionary regime it simultaneously warns against, on the apparent calculation that the goodwill buys faster approval for its own models and slower approval for Anthropic’s. That is not, by itself, an accusation of bad faith; it is a description of the incentives.
One can grant the strongest version of the case: the executive order is new, CAISI is new, and a case-by-case window may be the only available stopgap while statutory frameworks catch up. The argument collapses on contact with the record. The administration has used the same mechanism, in the same window, to shut Anthropic’s two most capable models out of foreign hands entirely — and out of U.S. hands by extension, since Anthropic halted all access to comply. That is not a stopgap. It is a cudgel. The “transition period” dressing is, in that context, not innocent — it is the language used to make cooperation look temporary when the architecture being built is durable.
The architecture is the part to keep your eye on. A case-by-case approval regime for the most capable AI models, run out of the executive branch with no statutory criteria and no judicial review, is not a security regime. It is a discretionary one, and the two are not the same. A security regime has rules; a discretionary regime has preferences. The executive order that gave cybersecurity and national-security officials a bigger role in evaluation can, by its own terms, be revised by the next executive order. CAISI’s role can be reduced, expanded, or abolished at the discretion of the same office that just used it to keep Anthropic’s customers offline. Nothing in the architecture as described prevents the next administration — or this one, eighteen months from now — from using the same mechanism for purposes that have nothing to do with biology or cybersecurity and everything to do with which companies, which countries, and which researchers are allowed to do frontier AI work at all.
The underlying technical capability is real, but that reality makes the informal vetting regime more illegitimate, not less. When a model demonstrates the ability to discover vulnerabilities in classified U.S. systems or complete autonomous cybersecurity tasks, the distribution of those weights is a public concern. The more consequential the capability, the more illegitimate it is to delegate its allocation to a discretionary, case-by-case Commerce Department handshake. The question is not whether the models should be secured; it is who gets to decide, and by what documented authority. The answer, at present, is a transition period of informal executive-branch vetting, where the firms that cooperate get to keep their models in the market, and the firms that do not get their access halted.
The frame for this kind of capture is that you do not get a security regime by adding discretion; you get one by removing it. The fix is not subtle. Published criteria that any company can read before submitting a model for review. An independent technical-evaluation body with statutory authority and fixed terms for its leadership. A default rule that capable models are released unless a specific, documented national-security finding supports withholding them. A private right of action for any company or researcher whose access is denied without such a finding. None of this requires new physics. It requires treating the distribution of frontier AI capability as the critical-infrastructure question it has, in the past two weeks, plainly become — and not as an executive-branch prerogative to be exercised case by case against whichever company has most recently fallen out of favor.
A formal charter would require an act of Congress, and neither side has any interest in that. It is the digital equivalent of a province deciding which mill gets to buy power from the hydro authority, not by a published tariff, but by a private negotiation with the minister. OpenAI is right that this should not become the long-term default. The trouble is that the regime OpenAI is cooperating with is the default — the architecture, not the weather. Until that architecture is changed, by statute, litigation, or a company willing to refuse to participate, the question of who gets to use the most capable AI in the world will be answered, case by case, by the same office that picked Anthropic to shut down and OpenAI to let through. The deadline for submissions on that question, as far as I can tell from here in Peterborough, has not been posted.