OpenAI Says It Doesn't Need an Antitrust Waiver. Two Cabinet-Level Officials Answered the Same Week.

On Tuesday 15 September 2026, OpenAI's global policy chief Chris Lehane told reporters that his company's engagement with Anthropic and Google DeepMind on AI safety had been under way for several weeks. He then volunteered something he had not been asked for: that the three largest frontier developers in the United States need no antitrust waiver to do it. His precedent was aviation. “There are many instances over time where companies are trying to help each other on safety,” he said, as reported by Bloomberg.

Two officials answered inside the same week. Andrew Ferguson, chair of the Federal Trade Commission, said the same day that he would be deeply suspicious of a request for such an exemption. At a House hearing, Treasury Secretary Scott Bessent told lawmakers not to grant the labs a liability exemption either. Neither was responding to a filed application, because there isn't one. They were responding to the claim.

Start with what is actually established. One executive at one company has characterised a set of conversations. Neither Anthropic nor Google DeepMind has published terms, a charter, a membership rule or an agenda. No joint statement exists. Everything below tests a legal theory, not a documented agreement — and the absence of documents is itself part of the problem the theory has.

The airline analogy, taken literally

Airlines do cooperate on safety, extensively, and almost none of it looks like three firms agreeing among themselves. The structure runs through the Federal Aviation Administration. The regulator sets airworthiness and operating standards; carriers report incidents into schemes the government designs and supervises; the analysis that comes out the other side is turned into directives that bind everyone, including operators who never sat in the room. Where voluntary reporting programmes exist, they exist because a federal agency created the channel and protects the reporter.

Three features do the legal work there, and each is worth naming because each is absent from the arrangement Lehane described. First, a government body, not the participants, decides what the standard is. Second, participation and disclosure obligations come from statute and rule rather than from a bargain between competitors. Third, the output is a requirement applied across the industry, not a shared view among the three largest firms about what any of them should ship.

Strip those out and what remains is the thing antitrust law was built to look at: the three companies with the largest frontier models talking to each other about the conditions under which frontier models get released.

What the enquiry actually turns on

Competitors collaborate lawfully all the time. Standards bodies are the ordinary case — USB, Wi-Fi, payment card security, video codecs. Antitrust does not treat standard-setting as suspect, and safety standards least of all. But the permission is conditional, and the conditions are structural rather than motivational.

The recurring conditions are open membership, terms offered on a fair and non-discriminatory basis, published procedures, and a clear line between agreeing on how something is measured and agreeing on what may be sold. A standards body that publishes a test is on solid ground. A group of dominant firms that agrees a rival's product should not ship until it passes their test is doing something else, whatever the test measures.

This is why a benign purpose has never been much of a defence on its own. Safety, quality and professional standards have all been offered as justifications for agreements that courts then examined for their effect on what competitors could offer. The question a regulator asks is not whether the participants meant well. It is whether the agreement, as structured, determines what someone else is allowed to bring to market.

Ferguson's stated objection is narrower and sharper than general suspicion, and it is worth quoting accurately because it names a mechanism. His position, as reported, is that the labs are seeking barriers to entry that insulate incumbents from challengers, and that the combination is what should trouble people: if companies are simultaneously asking for regulation and for an antitrust exemption, everyone should be deeply suspicious.

That pairing is the substance of the counter-argument. Regulation raises the fixed cost of competing. An antitrust exemption lets the firms that already carry that cost coordinate on what the regulation requires. Each half is defensible alone; a regulator's concern is what they do together.

“The best way to guarantee safety is that the creators are liable for what they build and generate. — Treasury Secretary Scott Bessent, testifying to House lawmakers, September 2026”

Hassabis already described the structure — and it wasn't three firms talking

The coordination Lehane referred to did not begin with OpenAI. In July 2026, Google DeepMind chief executive Demis Hassabis publicly called for a United States-led body to test frontier models for national security risk before release, covering models built domestically and abroad, open-weight and closed. Labs would share models with it voluntarily, up to thirty days ahead of release, for testing of dangerous cyber and biological capability and of whether safeguards can be bypassed.

His template was FINRA — the industry-funded self-regulatory organisation that polices broker-dealers under Securities and Exchange Commission oversight. That choice of analogy is more revealing than the airline one, because it concedes the point. FINRA is not an example of competitors coordinating without supervision. It is an example of an industry body whose rules require approval from a federal regulator before they bind anyone. The supervision is the reason the arrangement is lawful, and it is precisely the element a group of three firms talking among themselves does not have.

The proposal as described has no government approval and no binding authority, and it contemplated that the body could coordinate an industry-wide slowdown. A mechanism for agreeing when frontier releases pause is not information-sharing about threats. It is an agreement about output. Those are different objects in antitrust terms, and conflating them is what makes the waiver question live rather than academic.

Congress is being asked the question OpenAI says needs no asking

There is a legislative effort — reported as the Collaboration on Adversarial Threats and Security Risks Act — that would establish how antitrust law applies to the sharing of frontier model risk information among developers. DrafterDaily has not reviewed the bill text and is reporting its existence and stated purpose as described in coverage of the oversight-body debate.

The existence of such a bill is evidence in itself. Legislators do not spend floor time clarifying the applicability of antitrust law to conduct that is uncontroversially lawful. Somebody with counsel decided the question needed an answer in statute. Lehane may well be right that the easiest category — telling a rival that a specific jailbreak works, or that a particular capability showed up in evaluation — needs no exemption. That is close to pure information-sharing about a common threat, and it is the kind of exchange antitrust law has always had room for. The distance between that and a coordinated release pause is the whole dispute, and asserting the first does not settle the second.

What this does not establish

Three limits are worth stating plainly. The scope of any coordination rests on one executive's account; the other two participants have said nothing on the record about what is being discussed or agreed. No exemption has been requested, so Ferguson and Bessent were reacting to a public position rather than a filing. And the wider context that has been folded into this story is not evidence about it: Dario Amodei published a long essay on 12 September urging slower frontier development, and Sam Altman spoke publicly about slowdown on 14 September, but public advocacy by executives for the same policy is not an agreement between their companies.

There are two stable endings. One is a supervised body with published rules, open membership on non-discriminatory terms, and a statutory basis — roughly Hassabis's FINRA sketch, built properly, which is a slow process involving Congress. The other is an informal understanding among the three largest developers that works until a regulator or a smaller competitor asks to see it, at which point it is unwound. The second is cheaper to start and the reason to watch for the paperwork rather than the press availability.

The useful test is not whether the companies sound sincere about safety. It is whether the thing they build has a rulebook, a regulator, and a door that opens from the outside.