A TechTimes investigation published July 28 revealed that OpenAI, Anthropic, Google, Microsoft, and xAI are co-designing the federal threshold criteria that determine which frontier AI models face mandatory pre-release government scrutiny. The five labs writing the rules are the same five whose models have already cleared the system they are designing.

The framework sits inside Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” signed June 2, 2026. The order created a classified benchmarking process for covered frontier models, with the threshold definitions — the line that separates ordinary AI development from government-required safety review — being built in part by the companies most likely to be on both sides of that line.

What the Order Actually Does

EO 14409 requires frontier AI developers to notify the federal government before deploying models that exceed certain capability thresholds. What those thresholds are, exactly, is being determined through a process that includes direct input from the major labs.

The argument for this design is practical: the people building these models have the most relevant technical expertise. The argument against it is structural: the same companies defining the threshold are the ones whose existing models have already cleared it. Competitors who were not in the room when the definitions were written face a different set of conditions.

Both OpenAI and Anthropic have published competing visions of what federal AI regulation should look like. OpenAI’s blueprint, released June 3, calls for making the Commerce Department’s AI safety standards center the primary federal institution, building on laws already enacted in California, New York, and Illinois. Anthropic’s framework combines mandatory public safety requirements with independent evaluations and government authority to halt deployments that pose catastrophic risks, while opposing blanket federal preemption of state laws.

They agree on the threshold process. On how much power states should retain, they are on opposite sides.

The Pacing Letter

One day after the TechTimes investigation published, both companies formally endorsed what is being called the “Pacing the Frontier” letter. The statement, signed by 1,268 verified employees of frontier AI labs, calls on the US government to build international tools capable of deliberately slowing AI development if certain capability thresholds are crossed.

The specific trigger that has drawn the most attention in the letter is AI that can write its own training code. This is not a hypothetical concern in 2026. Multiple current AI systems demonstrate meaningful capability in this area. The signatories argue that once a model can meaningfully improve its own training, human oversight of the development process becomes qualitatively harder to maintain.

Both companies backing this letter while also being involved in designing the thresholds creates an unusual dynamic. They are simultaneously advocating for the government’s right to slow things down and helping determine where “fast enough to require slowing” begins.

What This Means for Business

For enterprise AI buyers and builders, there are three practical implications worth tracking.

The threshold is not public. Because the benchmarking criteria under EO 14409 are classified, neither enterprise buyers nor third-party researchers can independently verify whether a given model should have triggered government review before release. This creates opacity in the very part of the supply chain that compliance-sensitive businesses care most about.

Agentic AI tools are in scope. The Pacing letter’s specific focus on AI that writes its own code applies directly to tools like Claude Code, Codex, and similar systems that enterprises are adopting for software development. If those capabilities become restricted or access-controlled again, as happened temporarily with GPT-5.6 and Claude Fable 5 earlier this year, businesses built around those tools will face disruption.

Regulatory fragmentation is not resolving. OpenAI and Anthropic’s divergent positions on state versus federal preemption mean the patchwork of California, New York, Illinois, Colorado, and EU-level rules is likely to persist for the near term. Compliance planning needs to account for multiple jurisdictions, not a single federal standard.

The regulatory environment for enterprise AI is moving fast, in multiple directions at once. For businesses trying to build durable AI strategy rather than chasing announcements, the priority is governance frameworks that can adapt — not frameworks locked to a specific vendor, model, or set of rules that may look different in six months.

Enterprise DNA’s Omni Advisory service works with business leaders to build AI strategy and governance frameworks that hold up across changing regulatory conditions. Book a call to start the conversation.