Trump signs an AI executive order: government access to models 30 days before release — a turning point for US regulation

Trump signs an AI executive order: government access to models 30 days before release — a turning point for US regulation

🔎 Voluntary on paper, unavoidable in practice

On June 2, 2026, Donald Trump signed an executive order that could reshape the relationship between the federal government and AI giants. The principle: companies must provide their most powerful models to the government for evaluation before their public release. Except the keyword here is "voluntary." And that is precisely where it gets interesting.

Because behind this supposedly flexible framework lies an institutional architecture whose implications are much harsher than they appear. Between the intense lobbying by Musk and Zuckerberg, a one-month delay in the signing, and eleven days later a brutal blocking of Anthropic's models abroad, the June 2026 sequence tells a much more complex story than a simple "light EO".

The key points

• The June 2, 2026 EO asks AI companies for voluntary pre-release access to their advanced models for the federal government, for a review period limited to 30 days.
• This version is a stripped-down version of the initial draft, obtained after a one-month delay and intense lobbying by Elon Musk and Mark Zuckerberg, as reported by TechCrunch.
• The EO aims to preempt state laws on AI, but its legality is questionable according to NPR.
• On June 13, 2026, the government imposes export controls on Anthropic's Claude Mythos 5 and Fable 5, cutting off access to all foreign nationals — including Anthropic employees.
• This EO fits into a contradictory trajectory: cancellation of Biden's 2023 EO, followed by a pro-regulation U-turn in 2026.

Recommended Tools

From Biden to Trump: two radically opposed visions, one same objective

Biden's October 2023 EO imposed concrete obligations: mandatory reporting for models trained on more than 10^26 FLOPS, standardized safety tests, and sharing results with the government. It was a coercive framework, even if its actual scope remained limited.

Trump cancelled this EO as soon as he returned. Musk, Zuckerberg and Sacks then won a first battle: the end of direct federal constraints on AI development. The rhetoric was clear — innovation above all, no regulatory brakes.

Except that fifteen months later, Trump signs a new EO that reintroduces a government review mechanism. The difference? Everything is formally voluntary. But as Ropes & Gray analyzes, this voluntarism creates a "significant institutional architecture" whose de facto effects will be difficult for industry players to ignore.

The paradox is glaring: the same president who dismantled the Biden framework is rebuilding a supervisory device, just under another name.

What the EO actually says — decoding the text

The official text, published on the Maison Blanche website, establishes a framework of "voluntary pre-release review". Concretely, companies are invited to submit their advanced models to governmental "trusted partners".

The review is capped at 30 days. It does not concern all models, but only those deemed "advanced" — a threshold that remains deliberately vague in the text. The fact sheet officiel also recalls the context of the July 2025 EO prohibiting ideological biases in federal AI models, showing that this EO fits into a broader logic of political control of AI.

According to CNBC, the EO asks companies to give the government "early access" to their models. But the word "asks" makes all the difference compared to "requires".

This vagueness is calculated. It allows Trump to tell his base that he is not regulating AI, while giving the government a real lever of influence over the AI roadmap of the GAFAM.

Did Musk and Zuckerberg really "soften" the EO?

Yes, and sources are unanimous on this point. TechTimes reports that Musk and Zuckerberg actively lobbied to limit the scope of the EO. Their concrete victories: a 30-day cap on reviews, and restricting submissions to only "trusted partners" rather than the entire federal apparatus.

The one-month delay of the signing ceremony, confirmed by NPR, is directly linked to these industry objections. Trump did not want to sign a text that could have been perceived as anti-innovation — particularly in a context where competition with China over AI is a central political argument.

Musk, via xAI and Grok 4.1, and Zuckerberg, via Meta's investments in AI, had every interest in avoiding a process that could delay their launches. The result is a text that looks more like a gentlemen's club agreement than a regulation.

But you have to read between the lines. Even if "voluntary," a pre-release review mechanism creates an institutional precedent. Once the door is open, nothing prevents a future EO from making it mandatory.

Impact on OpenAI, Anthropic, and Google: who wins, who loses

Each player has a different position regarding this EO, and strategic calculations vary considerably.

OpenAI is probably the least impacted. With GPT-5.5 leading the agentic (98.2) and general (91) rankings, the company already has close channels with the US government. Sam Altman has publicly supported the idea of pre-release reviews since 2023. Participating in the framework strengthens its position as a privileged partner.

Google with Gemini 3.1 Pro (score 92 in general, 87.3 in agentic) finds itself in a similar position. The Mountain View giant has the resources to manage a 30-day review without significant impact on its roadmap. And the political credit with the administration is a non-negligible bonus.

Anthropic is the most fascinating and contradictory case. The company has built its identity on AI safety. Claude Opus 4.7 (Adaptive) reaches 94.3 in agentic and 90 in general. One might think that Anthropic would be the first supporter of a safety EO. Except that on June 13, 2026, eleven days after this "voluntary" EO, the government hit Anthropic with an export control that "abruptly disables" access to Fable 5 and Mythos 5 for any foreign national.

The irony is cruel. The company that advocated for caution finds itself with its most advanced models removed from the global market by a unilateral government decision.

The June 13 power play: Anthropic cut off from the world

On June 12, 2026, CNBC revealed that Anthropic was abruptly disabling access to its Mythos 5 and Fable 5 models. The official statement from Anthropic specifies that the suspension applies to any foreign national — including Anthropic's own employees based outside the United States.

Greek City Times confirms that this is a formal export control, literally blocking foreign access to these models. This is no longer "voluntary." It is a unilateral and coercive action.

The connection to the June 2 EO is troubling. This sequence can be read in two ways. Either the "voluntary" EO was a decoy, and the government never intended to settle for grace reviews. Or the two events are truly disconnected — the EO falls under homeland security, the export control under national security. But the timing coincidence is hard to ignore.

For Anthropic, the signal is clear: "voluntary" cooperation does not protect against unilateral state action. This is a lesson the entire industry should meditate on.

This tension between voluntary regulation and actual coercion also sheds light on the White House reversal on pre-release verification — what was presented as a gesture of trust is turning into a lever of control.

Voluntary vs mandatory: a distinction that isn't one

The firm Ropes & Gray published the most insightful legal analysis on this point. Their thesis: the voluntary nature of the EO creates "mandatory implications" through several indirect mechanisms.

First mechanism: reputational pressure. If OpenAI and Google participate and Anthropic or xAI refuse, the signal sent to the market and to Congress is negative. The "voluntary" becomes a test of good faith.

Second mechanism: conditioning federal contracts. Nothing prevents the government from favoring companies that participate in the scheme in its public procurement. The multi-vendor AI imposed on the Pentagon shows that the administration does not hesitate to use its purchasing power as a strategic lever.

Third mechanism: the legal basis for a future tightening. A voluntary EO creates the institutional infrastructure — the "trusted partners", the review protocols, the communication channels. Transforming this voluntary framework into an obligation would simply require a new EO. The infrastructure would already be in place.

This is exactly what happened with the export control on Anthropic: the government did not need a new legal framework, the Existing Export Administration Regulations were sufficient. But the June 2 EO normalized the idea that the government has a legitimate role in pre-release evaluation.

The war against state laws: the real motivation

According to NPR, one of the main goals of this EO was to preempt state AI laws. Several states, including California, Colorado, and New York, have advanced AI legislation proposals in recent months.

A federal EO, even voluntary, allows the administration to argue preemption — the idea that the federal government "occupies the field" and that states cannot legislate in parallel. This is a classic legal technique in the United States.

The problem is that the legality of this preemption is "questionable" according to NPR. A voluntary federal framework does not meet the usual conditions for preemption, which generally require complete regulatory occupation of the field. Future litigation before the Supreme Court is likely.

But in the meantime, the EO fulfills its political function: it gives the administration an argument to say "don't legislate, we're taking care of it." Even if "taking care of it" means a voluntary framework with significant de facto implications.

The broader context: ideological bias and political control

The White House fact sheet is not limited to technical security. It explicitly recalls the July 2025 EO prohibiting ideological bias in AI models used by the federal government.

This contextualizes the June 2026 EO within a broader logic of political control over AI. It is not only a question of whether a model is "safe" in the cybersecurity sense. It is also a question of whether it is "aligned" with the values of the incumbent administration.

This dimension is rarely discussed in technical analyses, but it is crucial. A pre-release review mechanism, even if voluntary, gives the executive branch early access to the models and the ability to exert pressure on their content. The fact that the "trusted partners" are not clearly identified in the text adds to the opacity.

For companies, the calculation is complex: participating in the framework means giving early access to the government, with all the political risks that this entails. Not participating means marginalizing oneself and potentially forfeiting federal markets.

Comparison of models affected by the EO

The EO does not precisely define the threshold for "advanced models," but in practice, it is the models at the top of the leaderboards that would be affected. Here is where things stand in June 2026.

General-purpose models potentially under review

Most impacted agentic models

The case of DeepSeek and Kimi K2.6 is revealing. These models would not be voluntarily submitted to the EO anyway — which highlights a fundamental limitation of the mechanism: it only applies to American companies. The competition with China, however invoked as a justification, is not addressed by this mechanism.

❌ Common mistakes

Mistake 1: Confusing "voluntary" with "without consequence"

The most widespread mistake is taking the word "voluntary" literally. As the Ropes & Gray analysis demonstrates, the framework creates institutional, reputational, and commercial pressures that make non-participation costly. Voluntarism is a tool of soft power, not a blank check.

Mistake 2: Thinking that the EO replaces the Biden EO

The 2026 EO does not restore the Biden framework. It creates a different mechanism — less coercitive on paper, but potentially more insidious because it relies on self-censorship rather than compliance. Companies under Biden had to fill out forms. Under Trump, they have to "choose" to cooperate. The difference in nature is important.

Mistake 3: Ignoring the export control on Anthropic

Many commentaries stopped at the June 2 EO without connecting it to the June 13 event. This is an analytical error. The sequence tells a coherent story: the government first establishes a "voluntary" framework, then uses its existing regulatory powers for coercive action. The context gives meaning to each individual event.

Mistake 4: Believing that the EO concerns all AI models

The text targets "advanced models," a deliberately vague term. In practice, a model like Claude Sonnet 4.6 (score 83 overall, 81.4 in agentic) might not be affected, whereas Claude Opus 4.7 would be. This asymmetry creates perverse incentives: companies might be tempted to limit the advertised capabilities of their models to stay below the threshold.

❓ Frequently Asked Questions

Is the June 2, 2026 EO mandatory for AI companies?

No. The text explicitly specifies that participation is voluntary. However, legal analyses emphasize that the de facto implications — reputational pressure, conditioning of federal contracts, creation of an institutional precedent — make non-participation strategically costly.

Why did Musk and Zuckerberg lobby against a stricter EO?

Because a mandatory pre-release review could have delayed their model launches and created regulatory uncertainties. Musk via xAI (Grok 4.1) and Zuckerberg via Meta's AI investments secured a 30-day cap and a restriction of reviews to "trusted partners" only.

What happened on June 13, 2026 for Anthropic?

The US government imposed an export control on Claude Mythos 5 and Fable 5, forcing Anthropic to disable access for any foreign national, including its own employees based outside the US. This is not formally linked to the June 2 EO, but the sequence is revealing.

How does this EO compare to the 2023 Biden EO?

The Biden EO imposed mandatory reporting and standardized testing. The Trump EO is formally voluntary and narrower. But it creates an institutional infrastructure that could facilitate a future tightening, and it serves as a pretext to block state laws.

Are Chinese companies like DeepSeek affected?

No. The voluntary EO only applies to companies under US jurisdiction. DeepSeek V4 Pro (Max) and Kimi K2.6 would not be affected, which highlights a structural limitation of the framework in the face of international competition.

✅ Conclusion

The June 2, 2026 EO is neither the long-awaited regulation nor the proclaimed laissez-faire — it is an institutional Trojan horse. Voluntary on the surface, it creates the conditions for growing government control over the roadmap of the most powerful AI models. The June 13 power grab against Anthropic shows that when "voluntary" is not enough, the state does not hesitate to turn to coercion. For developers and companies building on these APIs, the lesson is simple: US regulatory stability is an illusion. Prepare your multi-provider architectures — start by comparing the models available on Hostinger for your deployments, and diversify your dependencies before the next EO forces you to do so.