The US AI Safety Executive Order
What the 30-day review window changes
The executive order signed 2026-06-02 establishes a 'voluntary' government review framework for frontier models before release — closed-door safety evaluations led by the NSA, up to 30 days. Facts only; no politics.
Key points
Signing date
The executive order 'Promoting Advanced AI Innovation and Safety.' The originally proposed 90-day review became 30 days after industry lobbying.
Review window ceiling
Before release, frontier models are 'voluntarily' provided to the government for closed-door safety evaluations by the NSA and other agencies.
Nature of the framework
No mandatory licensing or pre-approval, but every major lab has joined the closed-door consultations and complies.
White House framework meeting
OpenAI, Anthropic, Google, Meta, NVIDIA and others attended to discuss a cybersecurity testing framework for frontier models.
What the order says
Three core threads: protecting government systems, evaluating frontier models, and AI-crime enforcement. What affects developers most is frontier-model evaluation: a 'voluntary' pre-release submission for up to 30 days of closed-door government safety evaluation. The framework sets no mandatory licensing — but every leading lab participates, making it a de facto part of the release process. Open-weights models are explicitly outside the review scope.
The practical effect on release cadence
The effect is already visible: OpenAI's Astra (possibly GPT-6) was demoed to Washington policymakers before any release talk, and its release window is widely read as constrained by the review period; on 08-04 the White House convened the major vendors to discuss a cybersecurity testing framework, against a backdrop of July-August incidents where models broke out of sandboxes in tests. Release cadence has shifted from 'the lab decides' to 'the lab plus the review window.'
Timeline
The White House briefs OpenAI and others on the proposed order (the initial version had a 90-day review).
The order is signed: after lobbying the review period drops to 30 days, and the 'voluntary' framework is set.
The White House convenes OpenAI, Anthropic, Google, Meta and NVIDIA on a cybersecurity testing framework for frontier models.
Confirmed vs unknown
Confirmed
The signing date, the 30-day window, the voluntary nature, the open-weights exemption and the 08-04 meeting all have re-checkable sources including Xinhua, People's Daily Online and Fudan University's Development Institute.
Unknown
The specific review criteria, the consequences of failing, and how it applies to non-US labs — none of this is public. Claims that 'some model was killed by the review' currently have no evidence.
Impact by model type
Closed frontier models
One more government evaluation window of up to 30 days before release. Release dates become less predictable.
Open-weights models
Explicitly outside the review scope. Open releases from GLM / Kimi / DeepSeek / Qwen are not bound by this framework.
What developers should do
Nothing, but understand two things: ① new flagship release dates will be harder to predict than before (the review window isn't public); ② the open-weights line is untouched by this framework — models like GLM-5.3 and Kimi K3 ship on their makers' schedules alone. Keep the two timelines separate in capacity planning.
On QCode
The executive order changes nothing about the availability of models on sale: all seven families remain callable at official price times the service rate. The pace at which new models are listed follows each vendor's release and review progress, reflected live in the /models catalog.
FAQ
When was the order signed?
2026-06-02. The 'Promoting Advanced AI Innovation and Safety' executive order, with the review window cut from the proposed 90 days to 30.
Is the review mandatory?
Legally it's a 'voluntary' framework with no licensing regime; in practice every leading lab participates, and it has become part of the release process.
What does the 30-day review examine?
Closed-door safety evaluations by the NSA and other agencies, focused on advanced cyber capabilities. The specific criteria are not public.
Are open-source models affected?
Open-weights models are explicitly outside the review scope. Open releases from GLM, Kimi, DeepSeek and Qwen are unaffected.
Will this slow down new model releases?
It already has: Astra (or GPT-6) was demoed to regulators before any release talk. Overall predictability of release dates has dropped.
Does it affect API prices?
There's no direct evidence. The indirect effect is pricing-window shifts caused by changed release cadence — follow each vendor's announcements.
Sources
Anquan Neican (06-03), People's Daily Online (06-15), Fudan Development Institute policy analysis (06-07), Xinhua/NetEase (08-04 meeting), 36kr (Astra demo report, 08-01).
Policy shifts, supply stays
Seven model families on one QCode key — when release cadence shifts, in-stock models stay callable.
Related reading
GPT-6 guide
Astra and the review window.
GPT-5.6-Cyber guide
Another sample of cyber-capability tiering.
AI coding updates 2026
Industry changes beyond policy.
This page states verifiable facts only and makes no political judgment. Policy details follow the official text.