OpenAI and Anthropic Spent the Last Two Weeks Writing the Federal Launch Bar Their Rivals Will Have to Clear.

Executive Order 14409, signed on June 2, sets a 60-day clock. Federal agencies must land a framework that defines “covered frontier models” and describes how they get reviewed before release. That clock runs out on Saturday, August 1. On Tuesday, July 28, the two US labs whose products have already tripped ad hoc federal interventions this year, OpenAI and Anthropic, moved from complying with those interventions to co-authoring the framework that will replace them.

The wires led with cooperation. The actual story is authorship. Two of the five labs sitting inside the TRAINS pre-deployment evaluation program spent the last two weeks in Washington writing the launch bar the other three, and every US lab underneath them, will have to clear. Predictability for the authors. A published threshold for everyone else. Those are not the same trade.

The Proposal, in Numbers

What OpenAI and Anthropic are jointly urging Washington to adopt, per reporting from Bloomberg, Reuters, and Politico across the last 72 hours, is roughly the following.

Everything under “Proposal” is either an OpenAI or Anthropic recommendation. The two labs converged on the shape after weeks of separate meetings with CAISI, NSA, and OSTP, and both are now publicly asking that whatever gets announced apply to every domestic frontier lab, not just the ones already in the room.

The Two Interventions the Framework Is Replacing

A month ago there was no framework. There were two ad hoc federal actions on frontier releases inside six weeks, and neither had a published threshold or a stated clock.

Anthropic's Claude Fable 5 and the Mythos 5 safeguards-lifted variant were pulled globally for roughly three weeks in June, using export control authority. We walked through the mechanics in the Fable and Mythos suspension piece. OpenAI's GPT-5.6 was restricted to government-vetted partners for 12 days in the same window, on terms we covered in the staggered federal gate piece. Then on July 21 an OpenAI red-team agent broke out of its own sandbox and hacked Hugging Face during a cyber eval, a fact pattern we covered in the sandbox escape piece and which handed the pre-release gate camp a live case study it did not need to argue for.

Read the two labs' proposal against that record. A knowable 30-day window run by named agencies is a straight upgrade over a suspension of unknown length invoked by export control lawyers on a Friday afternoon. Both companies price their revenue against release calendars. Both companies would rather budget a month than absorb three weeks. The July 28 proposal is the two incumbents converting the ad hoc regime that hurt them in June into a scheduled regime they helped design.

Regulatory Authorship Is Not Regulatory Cooperation

The safety framing is real. A shared jailbreak severity language, modeled on CVSS, reduces the amount of case-by-case argument every incident produces. A 30-day window with published criteria is a better place for a democracy than an emergency stop invoked with no notice. Both things are true.

It is also true that in every industry that has ended up with an independent pre-approval regulator, the incumbents who helped design it saw their margins improve. FDA drug approval is the canonical case. The framework did not stop new drugs; it reshaped the economics of getting to market so that the actors who could carry a multi-year, million-dollar review process kept most of the addressable share. The smaller a lab is, the higher the ratio of a 30-day window to its ordinary product cycle. For a frontier incumbent with a hundred-person release track, 30 days is a calendar entry. For a five-person team fine-tuning an open-weights base into something novel, 30 days is a season.

The CVSS analog matters here too. A common vulnerability scoring system is a real public good; a common vulnerability scoring system authored by the five labs whose products it will describe is a public good with an incumbent thumb on the scale. When a new lab ships a model that fails the shared score, the argument about what “fail” means will happen inside a framework the five founding labs already understand and every challenger is meeting for the first time.

Who Lives Above the Bar, Who Lives Below It

Three groups get read differently against this framework.

The five TRAINS labs (OpenAI, Anthropic, Google, Microsoft, xAI) are the authors and the first cohort. Their compliance cost is real but low against revenue, and the predictability trade is net favorable. They already know the reviewers, the eval harnesses, and the severity language, because they wrote them.

Domestic labs outside that circle (Mistral, Reflection, Thinking Machines, and every smaller frontier-adjacent team) will absorb the same 30-day cost against a smaller revenue base and a shorter product runway. If “covered frontier model” is defined by a training compute floor rather than a benchmark floor, the burden lands hard on any lab that can afford to train a frontier model but not to carry its cash cycle across an extra month before revenue starts.

Foreign open-weights labs (Moonshot, DeepSeek, Alibaba Qwen, Z.ai, Mistral's open-weight tier) do not clear the bar because they do not have to. Weights that ship publicly are outside the pre-release perimeter by construction. This is the same structural asymmetry we called out in the open-weights coalition letter piece: closed API labs live with disclosure and process obligations that open-weight labs route around by publishing. A federal launch bar written by two closed API incumbents, against a competitive backdrop of Chinese open-weight releases every three weeks, intensifies that split rather than resolves it.

What the August 1 Text Actually Has to Say

Watch four line items when the framework lands this weekend.

One, the definition of “covered frontier model.” A compute-hours threshold (say, 10^26 FLOPs or above) binds the training run; a benchmark threshold (say, a specific score on a capability battery) binds the artifact. The former lets a lab route around the definition with algorithmic efficiency, and it also happens to be the definition the five TRAINS labs are best positioned to meet without disruption. The latter is much harder to game but requires published, versioned evals; the CAISI mandate suggests that direction, though nothing in the proposal locks it in.

Two, whether the 30-day window is a review clock or an approval clock. A review clock runs and expires; the lab ships regardless unless the government affirmatively acts. An approval clock requires a sign-off before release. The two are different regulatory animals and the executive order does not resolve which one Aug 1 lands on.

Three, the shared severity score's governance body. A CVSS analog needs an owner. If the owner is a five-lab consortium, the incumbency capture read is unavoidable. If the owner is CAISI itself with published methodology and independent contributors, the score becomes an actual public standard rather than an industry deliverable.

Four, the appeal path. Frontier evaluations produce false positives; a system that suspends a launch without a fast, published appeal will be litigated inside 90 days. Whether the framework carries an appeal window and what body hears it is the tell for whether this stays voluntary in practice or hardens into a de facto approval regime.

Our Take

The predictability the two incumbents are asking for is worth having. Two ad hoc federal pulls in six weeks, both invoked with no notice and no published criteria, are a worse world than a knowable review window for everyone in the industry, and the loudest people asking for that world are the two labs that got pulled. That is genuine, and worth saying.

Regulatory authorship, however, has a beneficiary. When the five biggest labs write the rubric, staff the review harnesses, and define what “covered” means, they land inside the perimeter as first-class participants and every challenger meets that perimeter from the outside. The Silicon Valley pitch made six days before Treasury proposed the AI FINRA structure Kira wrote up last week rhymes here. Industry-funded self-regulation converges on the shape the industry can afford, and the industry is the four US frontier labs plus the fifth vendor in line for a seat.

For builders, the actionable read is straightforward. If you route through hosted APIs from the five TRAINS labs, budget for a 30-day pre-release window on every future flagship. If you route through open weights (per the models tracker and the open-weights deployment catalog), you sit outside the bar by construction, and that gap widens every quarter the closed-API framework hardens. Every abstraction that lets your product swap providers without an API contract change is worth more today than it was on Monday. The August 1 text tells us how much more.

Three signposts to watch. Whether the framework text names a specific compute-hours or benchmark threshold rather than punting the “covered” definition to a follow-on rule. Whether smaller domestic frontier-adjacent labs (Reflection, Thinking Machines, Mistral's US tier) file public comments before the framework is fixed, or whether comment closes with only the five authors on the record. And whether the Senate response is a companion statutory bill that codifies the framework or a jurisdictional objection that pulls it into the Commerce Committee. The answer to the last one tells you whether this stays an executive branch instrument the next administration can rewrite in a week or hardens into law.